Research ArticleA Novel Feature Extraction Technique Using Binarization ofBit Planes for Content Based Image Classification

Sudeep Thepade1 Rik Das2 and Saurav Ghosh3

1 Pimpri Chinchwad College of Engineering Akurdi Sector 26 Pradhikaran Nigdi Pune Maharashtra 411033 India2 Xavier Institute of Social Service Dr Camil Bulcke Path (Purulia Road) PO Box 7 Ranchi Jharkhand 834001 India3 AK Choudhury School of Information Technology University of Calcutta 92 APC Road Kolkata West Bengal 700009 India

Correspondence should be addressed to Rik Das rikdas78gmailcom

Received 8 July 2014 Revised 20 October 2014 Accepted 21 October 2014 Published 18 November 2014

Academic Editor Jie Zhou

Copyright copy 2014 SudeepThepade et alThis is an open access article distributed under theCreativeCommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

A number of techniques have been proposed earlier for feature extraction using image binarization Efficiency of the techniques wasdependent on proper threshold selection for the binarizationmethod In this paper a new feature extraction technique using imagebinarization has been proposed The technique has binarized the significant bit planes of an image by selecting local thresholdsThe proposed algorithm has been tested on a public dataset and has been compared with existing widely used techniques usingbinarization for extraction of features It has been inferred that the proposed method has outclassed all the existing techniques andhas shown consistent classification performance

1 Introduction

Incessant expansion of image datasets in terms of dimensionand complexity has escalated the requirement to designtechniques for efficient feature extraction Selection of imagefeatures has been the basis for content based image clas-sification as reviewed by Andreopoulos and Tsotsos in [1]In this work a new feature extraction technique applyingbinarization on bit planes using local threshold techniquehas been proposed A digital image can be separated intobit planes to understand the importance of each bit inthe image as shown by Thepade et al in [2] The processwas followed by binarization of significant bit planes forfeature vector extraction Binarization process calculated thethreshold value to differentiate the object of interest from itsbackgroundThe novel method has been compared quantita-tively with the techniques proposed by Thepade et al in [2]and by Kekre et al in [3] and four other widely used imagebinarization techniques proposed byNiblack [4] Bernsen [5]Sauvola and Pietikainen [6] and Otsu [7] Mean square error(MSE) method was followed for classification performance

evaluation of the proposed technique with respect to theexisting techniques for feature vector extraction

2 Related Work

Various methods have been used for feature extraction thathas implemented image binarization as a tool to denote theobject of interest and its background respectively Thresholdselection has been essential to facilitate binarization of imageto differentiate the object from its background Valizadehet al [8] Chang et al [9] and Gatos et al [10] havedescribed that threshold selection has been affected by anumber of factors including ambient illumination varianceof gray levels within the object and the background andinadequate contrast Process of threshold selection has beencategorized into three different techniques namely meanthreshold selection local threshold selection and globalthreshold selection Existing methods of feature extractionfrom images using selection of mean threshold were adoptedby Thepade et al in [2] and by Kekre et al in [3] The firstmethod of feature extraction using even and odd images [2]

Hindawi Publishing CorporationJournal of EngineeringVolume 2014 Article ID 439218 13 pageshttpdxdoiorg1011552014439218

2 Journal of Engineering

has generated two different varieties of images by addingand subtracting the original image and its flipped versionrespectively for each variety as shown in

119868even =(119868 + 119868

1015840)

2

119868odd =(119868 minus 119868

1015840)

2

(1)

where 119868 denotes the original image and 1198681015840 denotes its flippedversion

Binarization of even image and odd image were done bymean threshold technique as shown in (2) and the featurevectors of size 12 were generated

In the second case of feature extraction noteworthyinformation was extracted from the images by comparing thevalues of significant bits in each pixel and the extracted valueswere binarized following the process of mean thresholdselection for binarization [2] Equation (2) shows the processof mean threshold selection Consider

119879av119909 = (1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) (2)

where 119909 is red (119877) green (119866) and blue (119861) for each colorcomponent

Feature vectors of dimension 12 were generated by takingthe mean of significant values higher than the threshold andlower than the threshold respectively The third method forfeature extraction with multilevel block truncation coding[3] was an iterative process The initial stage started withbinarization using mean thresholding as in (2) and twofeature vectors for each color component were generatedfrom the mean of gray values higher than the thresholdand lower than the threshold respectively Classificationperformance was tested with the generated feature vectorsat the first level of feature vector extraction The next stepinvolved binarization of the gray values higher than themeanthreshold and lower than the mean threshold at the firstlevel This has produced four feature vectors for each colorcomponent and the classification performance was evaluatedwith the generated feature vectors at second level Now theclassification results for the first level of feature extractionand second level of feature extraction were compared to findout improvement in performance If the result for secondlevel was found to be higher than the first level then theentire feature extraction process was repeated for furtherlevels of feature extraction The process was continued untilthe classification performance deteriorated Multilevel blocktruncation coding method of feature extraction has shownits best performance at level 3 and the feature vector size atthe mentioned level was 24The drawback of mean thresholdmethod was to only determine a midpoint and has not beeneffective to differentiate the spread of data among the datasetsLocal threshold techniques discussed byNiblack [4] Bernsen[5] and Sauvola and Pietikainen [6] have used measures ofdispersion like standard deviation and variance to calculatethe threshold Both the local threshold selection techniques

proposed by Niblack Sauvola and Pietikainen have consid-ered the contribution of both mean and standard deviationfor threshold selection The methods involved sliding of a119887 times 119887 window over the image for calculation of thresholdvalue pixel-wise Sauvolarsquos technique was an improvementover Niblackrsquos method Bernsenrsquos technique of binarizationhas maintained a window size of 31 for threshold selectionBinarization process was performed locally by comparingeach pixel value to the corresponding threshold A pixel valuegreater than the corresponding threshold was represented by1 and a lower value than threshold was denoted by zero Thethreshold for each color component was computed as themean of the highest and the lowest gray value within thelocal window specified The contrast was expressed as thedistinction between the greatest and the least gray valuesBinarization was done by comparing the contrast value to thecontrast threshold Otsursquos method of binarization was basedon selection of global threshold The method categorized thepixels in an image into background and foreground pixelsby calculating the optimal threshold The drawback for thetechniques was the hefty feature vector size generated byeach The dimension of feature vectors was dependent onthe size of the image Thus for an image of size 119899 times 119899 thegenerated feature vector sizewas 1198992 Elalami [11] has extractedcolor and texture based features by 3D color histogram andGabor filters and addressed the problem for large featurevector size with genetic algorithm Hiremath and Pujari [12]have divided the image into nonoverlapping blocks Localdescriptors of color and texture were calculated from thecolor moments and moments on Gabor filter responses ofthese blocks were considered to generate local descriptorsof color and texture Gradient vector flow fields were usedto define the edge images for shape descriptors followedby the use of invariant moments Banerjee et al [13] havechosen visually significant point features from images andidentified the set by a fuzzy set theoretic approach Jalab[14] has combined color layout descriptor and Gabor texturedescriptor as a fusion approach for better identification ofimages Shen and Wu [15] have extracted feature vectorsfrom images by exploiting color texture and spatial structuredescriptors Irtaza et al [16] have used wavelet packets andEigen values of Gabor filters for creating feature vectorsfrom images Rahimi and Moghaddam [17] have exploredthe intraclass and interclass features for effective extractionof image signatures The proposed technique was comparedto seven other state-of-the-art feature extraction techniquesfor assessment of classification performanceThe quantitativeevaluation and statistical analyses have revealed the efficiencyof the proposed method

3 Proposed Methodology

The proposed method in this work can be subdivided intotwo different levels namely Bit Plane Slicing and thresholdselection as given below

31 Bit Plane Slicing Individual bits needed to represent anumber in binary have been considered as the bit planes

Journal of Engineering 3

8 bit binary253

1significant bit

1111101

significant bit

Bit-8 most

Bit-2Bit-1 least

Figure 1 Bit planes represented by 8 bit binary vector

The first level started with segmentation of the images intovarious bit planes starting from the least significant bit (LSB)to the most significant bit (MSB) The intensity value of eachpixel was denoted by an 8-bit binary vector and the value waseither 0 or 1 as shown in Figure 1 Thus each bit plane wasrepresented by a binary matrix which was further used togenerate image slices for respective bit planes as in Figure 2Significant bit planes carrying rich information were identi-fied and considered for feature extraction by binarization asin (3) and insignificant bit planes were discarded Consider

119909bp = 119909bp(bp5=1sdotsdotsdotorsdotsdotsdotbp6=1sdotsdotsdotorsdotsdotsdotbp7=1sdotsdotsdotorsdotsdotsdotbp8=1) (3)

where119909 is119877119866 and119861 respectively for each color component

32 Threshold Selection The test dataset contains differentscenes from diverse categories having varying illuminationacross the scene as observed in Figure 7 Uneven luminancedistribution caused the image to be brighter or darker atplaces which has adversely affected the recognition of theobject of interest Thus choosing a single global or meanthreshold for the entire scenes of diverse categories was notan effective alternative Instead selection of local thresholdtechnique can handle the problem of uneven illuminationby selecting threshold locally A popular local thresholdselection method named Niblackrsquos method [4] was used forbinarization of significant bit planes of the images fromdifferent categories Pixel-wise threshold was calculated bythe method for each color component of the selected bitplanes by sliding a rectangular window over the componentThe threshold was determined based on the local mean119898(119894 119895) and standard deviation 120590 (119894 119895) The window size wasconsidered as (25 times 25) The threshold was given as 119879 (119894 119895) =119898 (119894 119895) + 119896 sdot 120590 (119894 119895) Here 119896 has been a constant havingvalue between 0 and 1 and was considered to be 06 in theproposed method The value of 119896 and the size of slidingwindow determined the quality of binarization Figure 3(a)has shown the original image fromwhich bit plane 5 has beenextracted in Figure 3(b) Figure 3(c) has shown the effect ofbinarization on the original image usingmean thresholdTheimage of bit plane 5 has been binarized with mean thresholdin Figure 3(d) Finally Niblackrsquos method of local thresholdwith 119896 = 06 and window size of 25 times 25 has been used tobinarize the image of bit plane 5 in Figure 3(e) The effect

of binarization with mean threshold and Niblackrsquos methodof local threshold has been shown with different bit planessimilarly from Figures 4 5 and 6

33 Proposed Algorithm

Begin

(1) Input an image 119868 with three different colorcomponents 119877 119866 and 119861 respectively of size119898 lowast 119899 each

(2) Calculate the local threshold value 119879119909 for eachpixel in each color component119877119866 and 119861 usingNiblackrsquos method Consider

119879119909 = (1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) + 119896 lowast stddev (4)

where 119896 = 06lowast119909 = 119877 119866 and 119861 lowast

(3) Compute binary image maps for each pixel forthe given image Consider

Bitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 1198791199090 if 119909 (119894 119895) lt 119879119909

(5)

lowast119909 = 119877 119866 and 119861 lowast(4) Generate image features for the given image for

each color component Consider

119883upmean = 1

sum119898

119894=1sum119899

119895=1BitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BitMap119909(119894 119895) lowast 119883 (119894 119895)

(6)

119883lomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1Bitmap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BitMap119909(119894 119895)) lowast 119883 (119894 119895)

(7)

lowast119909 = 119877 119866 and 119861 lowast(5) Identify the higher bit planes for each color

component starting from bit plane 5 to bit plane8 which is equal to 1 Consider

119909bp = 119909bp(bp5=1sdotsdotsdotorsdotsdotsdotbp6=1sdotsdotsdotorsdotsdotsdotbp7=1sdotsdotsdotorsdotsdotsdotbp8=1) (8)

lowast119909 = 119877 119866 and 119861 lowast(6) Compute local threshold for the identified sig-

nificant bit planes for each color componentusing Niblackrsquos method Consider

119879119909bp = ( 1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) + 119896 lowast stddev (9)

where 119896 = 06lowast119909 = 119877 119866 and 119861 lowast

4 Journal of Engineering

(a) (b) (c)

(d) (e) (f)

Figure 2 (a) Original image (b) image of bit plane 5 (c) image of bit plane 6 (d) image of bit plane 7 (e) image of bit plane 8 and (f)amalgamated image of bit planes 5 6 7 and 8

(a) (b) (c) (d) (e)

Figure 3 (a) Original image (b) image of bit plane 5 (c) binarization of original image using mean threshold (d) binarization of bit plane 5with mean threshold and (e) binarization of bit plane 5 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 4 (a) Original image (b) image of bit plane 6 (c) binarization of original image using mean threshold (d) binarization of bit plane6 with mean threshold and (e) binarization of bit plane 6 with Niblackrsquos method of local threshold

Journal of Engineering 5

(a) (b) (c) (d) (e)

Figure 5 (a) Original image (b) image of bit plane 7 (c) binarization of original image using mean threshold (d) binarization of bit plane 7with mean threshold and (e) binarization of bit plane 7 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 6 (a) Original image (b) image of bit plane 8 (c) binarization of original image using mean threshold (d) binarization of bit plane8 with mean threshold and (e) binarization of bit plane 8 with Niblackrsquos method of local threshold

(7) Compute binary image maps for each pixel ofsignificant bit planes for 119877 119866 and 119861 respec-tively Consider

Bpbitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 119879119909bp0 if 119909 (119894 119895) lt 119879119909bp

(10)

lowast119909 = 119877 119866 and 119861 lowast(8) Compute the feature vectors 119909bitplanehi and

119909bitplanelo for bit planes for all the three colorcomponents Consider

119883bpupmean = 1

sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BPBitMap119909(119894 119895) lowast 119883 (119894 119895)

(11)

119883bplomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BPBitMap119909(119894 119895)) lowast 119883 (119894 119895)

(12)

lowast119909 = 119877 119866 and 119861 lowast(9) The feature vectors 119883upmean 119883lomean

119883bpupmean and 119883bplomean for each colorcomponent are associated to form twelvefeature vectors altogether for each image in thedataset

End

4 Experimental Process

The proposed technique was experimented with a subsetof Corel stock photo database known as Wang datasetused by Li and Wang in [18] It has been considered as awidely used public dataset The experimental process hasnot compromised with the size and quality of the imagesThe dataset comprised 9 categories with 100 images in eachcategory Figure 7 shows a sample of the original databaseThe classification performances were evaluated with 10-foldcross-validation scheme for the proposed and existing featurevector extraction techniques The process has been called 119899-fold cross-validation as given by Sridhar in [19] The valueof 119899 is an integer and was considered to be 10 in this workThe entire dataset was divided into 10 subsets 1 subset wasconsidered as the testing set and the remaining 9 subsets wereconsidered to be training sets The method was repeated for10 trials and the performance of the classifiers was evaluatedby combining the 10 results thus obtained after evaluating the10 folds

5 Evaluation of Proposed Technique

The proposed technique for feature extraction has beenevaluated by mean square error (MSE) method as in (13)The method considered the similarity measure between twoinstances as described by Xu et al in [20] and Kotsiantisin [21] The nearest neighbor in the instance space waslocated for classification and then the unknown instancewas designated with the same class of the identified nearestneighbor Consider

MSE = 1

119872119873

119872

sum

119910=1

119873

sum

119909=1

[119868 (119909 119910) minus 1198681015840(119909 119910)]

2

(13)

6 Journal of Engineering

Table 1 Comparison of misclassification rate (MR) for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses MountainsMisclassification rate (MR) 008 011 015 007 000 006 002 002 012

Table 2 Comparison of 1198651 score for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses Mountains1198651 score 064 050 038 070 100 072 091 093 042

Table 3 Confusion matrix

Cat1 Cat2 Cat3 Cat4 Cat5 Cat6 Cat7 Cat8 Cat9 Classified as61 2 15 11 1 8 0 2 0 Tribals3 51 12 3 0 8 1 0 22 Sea beach9 10 41 15 0 4 4 0 17 Gothic structure8 3 13 71 0 0 0 0 5 Bus0 0 0 0 100 0 0 0 0 Dinosaur6 9 8 0 0 72 0 2 3 Elephant2 1 3 0 0 1 89 4 0 Roses0 2 1 0 0 2 0 95 0 Horses1 26 23 4 0 5 1 1 39 Mountains

Figure 7 Sample images of different categories from Wangdatabase

where 119868 and 1198681015840 are the two images used for comparison usingthe MSE method

Primarily two different evaluation metrics namely mis-classification rate (MR) and 1198651 score were considered tocompare the proposed feature extraction technique withrespect to the existing techniques

51MisclassificationRate (MR) Theerror rate of the classifierindicates the proportion of instances that have been wronglyclassified as in

MR =FP + FN

TP + TN + FP + FN (14)

where true positive (TP) is number of instances classifiedcorrectly true negative (TN) is number of negative resultscreated for negative instances false positive (FP) is numberof erroneous results as positive results for negative instancesand false negative (FN) is number of erroneous results asnegative results for positive instances

52 F1-Score Precision andRecall (TP rate) can be combinedto produce a metric known as 1198651 score as in (15) It has beenconsidered as the harmonic mean of Precision and RecallHigher value of1198651 score indicates better classification resultsConsider

1198651score = 2 lowast Precision lowast RecallPrecision + Recall

(15)

where Precision is the probability that an object is classifiedcorrectly as per the actual value and Recall is the probabilityof a classifier that it will produce true positive result

The comparison shown in Tables 1 and 2 has beengraphically represented in Figures 8 and 9 It has beenobserved that the images of category named dinosaur havegot minimummisclassification rate and the highest 1198651 scoreOn the other hand least classification performance has beenexhibited by the category named gothic structure as shown inFigures 8 and 9The category named gothic structure has thehighest misclassification rate (MR) and the lowest 1198651 scorewhich signified poor feature extraction compared to the othercategories in the dataset The confusion matrix for all thecategories has been given as in Table 3

6 Results and Discussion

The proposed technique has implemented feature extractionby binarization of significant bit planes with Niblackrsquos local

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

2 Journal of Engineering

has generated two different varieties of images by addingand subtracting the original image and its flipped versionrespectively for each variety as shown in

119868even =(119868 + 119868

1015840)

2

119868odd =(119868 minus 119868

1015840)

2

(1)

where 119868 denotes the original image and 1198681015840 denotes its flippedversion

Binarization of even image and odd image were done bymean threshold technique as shown in (2) and the featurevectors of size 12 were generated

In the second case of feature extraction noteworthyinformation was extracted from the images by comparing thevalues of significant bits in each pixel and the extracted valueswere binarized following the process of mean thresholdselection for binarization [2] Equation (2) shows the processof mean threshold selection Consider

119879av119909 = (1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) (2)

where 119909 is red (119877) green (119866) and blue (119861) for each colorcomponent

Feature vectors of dimension 12 were generated by takingthe mean of significant values higher than the threshold andlower than the threshold respectively The third method forfeature extraction with multilevel block truncation coding[3] was an iterative process The initial stage started withbinarization using mean thresholding as in (2) and twofeature vectors for each color component were generatedfrom the mean of gray values higher than the thresholdand lower than the threshold respectively Classificationperformance was tested with the generated feature vectorsat the first level of feature vector extraction The next stepinvolved binarization of the gray values higher than themeanthreshold and lower than the mean threshold at the firstlevel This has produced four feature vectors for each colorcomponent and the classification performance was evaluatedwith the generated feature vectors at second level Now theclassification results for the first level of feature extractionand second level of feature extraction were compared to findout improvement in performance If the result for secondlevel was found to be higher than the first level then theentire feature extraction process was repeated for furtherlevels of feature extraction The process was continued untilthe classification performance deteriorated Multilevel blocktruncation coding method of feature extraction has shownits best performance at level 3 and the feature vector size atthe mentioned level was 24The drawback of mean thresholdmethod was to only determine a midpoint and has not beeneffective to differentiate the spread of data among the datasetsLocal threshold techniques discussed byNiblack [4] Bernsen[5] and Sauvola and Pietikainen [6] have used measures ofdispersion like standard deviation and variance to calculatethe threshold Both the local threshold selection techniques

proposed by Niblack Sauvola and Pietikainen have consid-ered the contribution of both mean and standard deviationfor threshold selection The methods involved sliding of a119887 times 119887 window over the image for calculation of thresholdvalue pixel-wise Sauvolarsquos technique was an improvementover Niblackrsquos method Bernsenrsquos technique of binarizationhas maintained a window size of 31 for threshold selectionBinarization process was performed locally by comparingeach pixel value to the corresponding threshold A pixel valuegreater than the corresponding threshold was represented by1 and a lower value than threshold was denoted by zero Thethreshold for each color component was computed as themean of the highest and the lowest gray value within thelocal window specified The contrast was expressed as thedistinction between the greatest and the least gray valuesBinarization was done by comparing the contrast value to thecontrast threshold Otsursquos method of binarization was basedon selection of global threshold The method categorized thepixels in an image into background and foreground pixelsby calculating the optimal threshold The drawback for thetechniques was the hefty feature vector size generated byeach The dimension of feature vectors was dependent onthe size of the image Thus for an image of size 119899 times 119899 thegenerated feature vector sizewas 1198992 Elalami [11] has extractedcolor and texture based features by 3D color histogram andGabor filters and addressed the problem for large featurevector size with genetic algorithm Hiremath and Pujari [12]have divided the image into nonoverlapping blocks Localdescriptors of color and texture were calculated from thecolor moments and moments on Gabor filter responses ofthese blocks were considered to generate local descriptorsof color and texture Gradient vector flow fields were usedto define the edge images for shape descriptors followedby the use of invariant moments Banerjee et al [13] havechosen visually significant point features from images andidentified the set by a fuzzy set theoretic approach Jalab[14] has combined color layout descriptor and Gabor texturedescriptor as a fusion approach for better identification ofimages Shen and Wu [15] have extracted feature vectorsfrom images by exploiting color texture and spatial structuredescriptors Irtaza et al [16] have used wavelet packets andEigen values of Gabor filters for creating feature vectorsfrom images Rahimi and Moghaddam [17] have exploredthe intraclass and interclass features for effective extractionof image signatures The proposed technique was comparedto seven other state-of-the-art feature extraction techniquesfor assessment of classification performanceThe quantitativeevaluation and statistical analyses have revealed the efficiencyof the proposed method

3 Proposed Methodology

The proposed method in this work can be subdivided intotwo different levels namely Bit Plane Slicing and thresholdselection as given below

31 Bit Plane Slicing Individual bits needed to represent anumber in binary have been considered as the bit planes

Journal of Engineering 3

8 bit binary253

1significant bit

1111101

significant bit

Bit-8 most

Bit-2Bit-1 least

Figure 1 Bit planes represented by 8 bit binary vector

The first level started with segmentation of the images intovarious bit planes starting from the least significant bit (LSB)to the most significant bit (MSB) The intensity value of eachpixel was denoted by an 8-bit binary vector and the value waseither 0 or 1 as shown in Figure 1 Thus each bit plane wasrepresented by a binary matrix which was further used togenerate image slices for respective bit planes as in Figure 2Significant bit planes carrying rich information were identi-fied and considered for feature extraction by binarization asin (3) and insignificant bit planes were discarded Consider

119909bp = 119909bp(bp5=1sdotsdotsdotorsdotsdotsdotbp6=1sdotsdotsdotorsdotsdotsdotbp7=1sdotsdotsdotorsdotsdotsdotbp8=1) (3)

where119909 is119877119866 and119861 respectively for each color component

32 Threshold Selection The test dataset contains differentscenes from diverse categories having varying illuminationacross the scene as observed in Figure 7 Uneven luminancedistribution caused the image to be brighter or darker atplaces which has adversely affected the recognition of theobject of interest Thus choosing a single global or meanthreshold for the entire scenes of diverse categories was notan effective alternative Instead selection of local thresholdtechnique can handle the problem of uneven illuminationby selecting threshold locally A popular local thresholdselection method named Niblackrsquos method [4] was used forbinarization of significant bit planes of the images fromdifferent categories Pixel-wise threshold was calculated bythe method for each color component of the selected bitplanes by sliding a rectangular window over the componentThe threshold was determined based on the local mean119898(119894 119895) and standard deviation 120590 (119894 119895) The window size wasconsidered as (25 times 25) The threshold was given as 119879 (119894 119895) =119898 (119894 119895) + 119896 sdot 120590 (119894 119895) Here 119896 has been a constant havingvalue between 0 and 1 and was considered to be 06 in theproposed method The value of 119896 and the size of slidingwindow determined the quality of binarization Figure 3(a)has shown the original image fromwhich bit plane 5 has beenextracted in Figure 3(b) Figure 3(c) has shown the effect ofbinarization on the original image usingmean thresholdTheimage of bit plane 5 has been binarized with mean thresholdin Figure 3(d) Finally Niblackrsquos method of local thresholdwith 119896 = 06 and window size of 25 times 25 has been used tobinarize the image of bit plane 5 in Figure 3(e) The effect

of binarization with mean threshold and Niblackrsquos methodof local threshold has been shown with different bit planessimilarly from Figures 4 5 and 6

33 Proposed Algorithm

Begin

(1) Input an image 119868 with three different colorcomponents 119877 119866 and 119861 respectively of size119898 lowast 119899 each

(2) Calculate the local threshold value 119879119909 for eachpixel in each color component119877119866 and 119861 usingNiblackrsquos method Consider

119879119909 = (1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) + 119896 lowast stddev (4)

where 119896 = 06lowast119909 = 119877 119866 and 119861 lowast

(3) Compute binary image maps for each pixel forthe given image Consider

Bitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 1198791199090 if 119909 (119894 119895) lt 119879119909

(5)

lowast119909 = 119877 119866 and 119861 lowast(4) Generate image features for the given image for

each color component Consider

119883upmean = 1

sum119898

119894=1sum119899

119895=1BitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BitMap119909(119894 119895) lowast 119883 (119894 119895)

(6)

119883lomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1Bitmap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BitMap119909(119894 119895)) lowast 119883 (119894 119895)

(7)

lowast119909 = 119877 119866 and 119861 lowast(5) Identify the higher bit planes for each color

component starting from bit plane 5 to bit plane8 which is equal to 1 Consider

119909bp = 119909bp(bp5=1sdotsdotsdotorsdotsdotsdotbp6=1sdotsdotsdotorsdotsdotsdotbp7=1sdotsdotsdotorsdotsdotsdotbp8=1) (8)

lowast119909 = 119877 119866 and 119861 lowast(6) Compute local threshold for the identified sig-

nificant bit planes for each color componentusing Niblackrsquos method Consider

119879119909bp = ( 1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) + 119896 lowast stddev (9)

where 119896 = 06lowast119909 = 119877 119866 and 119861 lowast

4 Journal of Engineering

(a) (b) (c)

(d) (e) (f)

Figure 2 (a) Original image (b) image of bit plane 5 (c) image of bit plane 6 (d) image of bit plane 7 (e) image of bit plane 8 and (f)amalgamated image of bit planes 5 6 7 and 8

(a) (b) (c) (d) (e)

Figure 3 (a) Original image (b) image of bit plane 5 (c) binarization of original image using mean threshold (d) binarization of bit plane 5with mean threshold and (e) binarization of bit plane 5 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 4 (a) Original image (b) image of bit plane 6 (c) binarization of original image using mean threshold (d) binarization of bit plane6 with mean threshold and (e) binarization of bit plane 6 with Niblackrsquos method of local threshold

Journal of Engineering 5

(a) (b) (c) (d) (e)

Figure 5 (a) Original image (b) image of bit plane 7 (c) binarization of original image using mean threshold (d) binarization of bit plane 7with mean threshold and (e) binarization of bit plane 7 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 6 (a) Original image (b) image of bit plane 8 (c) binarization of original image using mean threshold (d) binarization of bit plane8 with mean threshold and (e) binarization of bit plane 8 with Niblackrsquos method of local threshold

(7) Compute binary image maps for each pixel ofsignificant bit planes for 119877 119866 and 119861 respec-tively Consider

Bpbitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 119879119909bp0 if 119909 (119894 119895) lt 119879119909bp

(10)

lowast119909 = 119877 119866 and 119861 lowast(8) Compute the feature vectors 119909bitplanehi and

119909bitplanelo for bit planes for all the three colorcomponents Consider

119883bpupmean = 1

sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BPBitMap119909(119894 119895) lowast 119883 (119894 119895)

(11)

119883bplomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BPBitMap119909(119894 119895)) lowast 119883 (119894 119895)

(12)

lowast119909 = 119877 119866 and 119861 lowast(9) The feature vectors 119883upmean 119883lomean

119883bpupmean and 119883bplomean for each colorcomponent are associated to form twelvefeature vectors altogether for each image in thedataset

End

4 Experimental Process

The proposed technique was experimented with a subsetof Corel stock photo database known as Wang datasetused by Li and Wang in [18] It has been considered as awidely used public dataset The experimental process hasnot compromised with the size and quality of the imagesThe dataset comprised 9 categories with 100 images in eachcategory Figure 7 shows a sample of the original databaseThe classification performances were evaluated with 10-foldcross-validation scheme for the proposed and existing featurevector extraction techniques The process has been called 119899-fold cross-validation as given by Sridhar in [19] The valueof 119899 is an integer and was considered to be 10 in this workThe entire dataset was divided into 10 subsets 1 subset wasconsidered as the testing set and the remaining 9 subsets wereconsidered to be training sets The method was repeated for10 trials and the performance of the classifiers was evaluatedby combining the 10 results thus obtained after evaluating the10 folds

5 Evaluation of Proposed Technique

The proposed technique for feature extraction has beenevaluated by mean square error (MSE) method as in (13)The method considered the similarity measure between twoinstances as described by Xu et al in [20] and Kotsiantisin [21] The nearest neighbor in the instance space waslocated for classification and then the unknown instancewas designated with the same class of the identified nearestneighbor Consider

MSE = 1

119872119873

119872

sum

119910=1

119873

sum

119909=1

[119868 (119909 119910) minus 1198681015840(119909 119910)]

2

(13)

6 Journal of Engineering

Table 1 Comparison of misclassification rate (MR) for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses MountainsMisclassification rate (MR) 008 011 015 007 000 006 002 002 012

Table 2 Comparison of 1198651 score for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses Mountains1198651 score 064 050 038 070 100 072 091 093 042

Table 3 Confusion matrix

Cat1 Cat2 Cat3 Cat4 Cat5 Cat6 Cat7 Cat8 Cat9 Classified as61 2 15 11 1 8 0 2 0 Tribals3 51 12 3 0 8 1 0 22 Sea beach9 10 41 15 0 4 4 0 17 Gothic structure8 3 13 71 0 0 0 0 5 Bus0 0 0 0 100 0 0 0 0 Dinosaur6 9 8 0 0 72 0 2 3 Elephant2 1 3 0 0 1 89 4 0 Roses0 2 1 0 0 2 0 95 0 Horses1 26 23 4 0 5 1 1 39 Mountains

Figure 7 Sample images of different categories from Wangdatabase

where 119868 and 1198681015840 are the two images used for comparison usingthe MSE method

Primarily two different evaluation metrics namely mis-classification rate (MR) and 1198651 score were considered tocompare the proposed feature extraction technique withrespect to the existing techniques

51MisclassificationRate (MR) Theerror rate of the classifierindicates the proportion of instances that have been wronglyclassified as in

MR =FP + FN

TP + TN + FP + FN (14)

where true positive (TP) is number of instances classifiedcorrectly true negative (TN) is number of negative resultscreated for negative instances false positive (FP) is numberof erroneous results as positive results for negative instancesand false negative (FN) is number of erroneous results asnegative results for positive instances

52 F1-Score Precision andRecall (TP rate) can be combinedto produce a metric known as 1198651 score as in (15) It has beenconsidered as the harmonic mean of Precision and RecallHigher value of1198651 score indicates better classification resultsConsider

1198651score = 2 lowast Precision lowast RecallPrecision + Recall

(15)

where Precision is the probability that an object is classifiedcorrectly as per the actual value and Recall is the probabilityof a classifier that it will produce true positive result

The comparison shown in Tables 1 and 2 has beengraphically represented in Figures 8 and 9 It has beenobserved that the images of category named dinosaur havegot minimummisclassification rate and the highest 1198651 scoreOn the other hand least classification performance has beenexhibited by the category named gothic structure as shown inFigures 8 and 9The category named gothic structure has thehighest misclassification rate (MR) and the lowest 1198651 scorewhich signified poor feature extraction compared to the othercategories in the dataset The confusion matrix for all thecategories has been given as in Table 3

6 Results and Discussion

The proposed technique has implemented feature extractionby binarization of significant bit planes with Niblackrsquos local

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Engineering 3

8 bit binary253

1significant bit

1111101

significant bit

Bit-8 most

Bit-2Bit-1 least

Figure 1 Bit planes represented by 8 bit binary vector

The first level started with segmentation of the images intovarious bit planes starting from the least significant bit (LSB)to the most significant bit (MSB) The intensity value of eachpixel was denoted by an 8-bit binary vector and the value waseither 0 or 1 as shown in Figure 1 Thus each bit plane wasrepresented by a binary matrix which was further used togenerate image slices for respective bit planes as in Figure 2Significant bit planes carrying rich information were identi-fied and considered for feature extraction by binarization asin (3) and insignificant bit planes were discarded Consider

119909bp = 119909bp(bp5=1sdotsdotsdotorsdotsdotsdotbp6=1sdotsdotsdotorsdotsdotsdotbp7=1sdotsdotsdotorsdotsdotsdotbp8=1) (3)

where119909 is119877119866 and119861 respectively for each color component

32 Threshold Selection The test dataset contains differentscenes from diverse categories having varying illuminationacross the scene as observed in Figure 7 Uneven luminancedistribution caused the image to be brighter or darker atplaces which has adversely affected the recognition of theobject of interest Thus choosing a single global or meanthreshold for the entire scenes of diverse categories was notan effective alternative Instead selection of local thresholdtechnique can handle the problem of uneven illuminationby selecting threshold locally A popular local thresholdselection method named Niblackrsquos method [4] was used forbinarization of significant bit planes of the images fromdifferent categories Pixel-wise threshold was calculated bythe method for each color component of the selected bitplanes by sliding a rectangular window over the componentThe threshold was determined based on the local mean119898(119894 119895) and standard deviation 120590 (119894 119895) The window size wasconsidered as (25 times 25) The threshold was given as 119879 (119894 119895) =119898 (119894 119895) + 119896 sdot 120590 (119894 119895) Here 119896 has been a constant havingvalue between 0 and 1 and was considered to be 06 in theproposed method The value of 119896 and the size of slidingwindow determined the quality of binarization Figure 3(a)has shown the original image fromwhich bit plane 5 has beenextracted in Figure 3(b) Figure 3(c) has shown the effect ofbinarization on the original image usingmean thresholdTheimage of bit plane 5 has been binarized with mean thresholdin Figure 3(d) Finally Niblackrsquos method of local thresholdwith 119896 = 06 and window size of 25 times 25 has been used tobinarize the image of bit plane 5 in Figure 3(e) The effect

of binarization with mean threshold and Niblackrsquos methodof local threshold has been shown with different bit planessimilarly from Figures 4 5 and 6

33 Proposed Algorithm

Begin

(1) Input an image 119868 with three different colorcomponents 119877 119866 and 119861 respectively of size119898 lowast 119899 each

(2) Calculate the local threshold value 119879119909 for eachpixel in each color component119877119866 and 119861 usingNiblackrsquos method Consider

119879119909 = (1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) + 119896 lowast stddev (4)

where 119896 = 06lowast119909 = 119877 119866 and 119861 lowast

(3) Compute binary image maps for each pixel forthe given image Consider

Bitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 1198791199090 if 119909 (119894 119895) lt 119879119909

(5)

lowast119909 = 119877 119866 and 119861 lowast(4) Generate image features for the given image for

each color component Consider

119883upmean = 1

sum119898

119894=1sum119899

119895=1BitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BitMap119909(119894 119895) lowast 119883 (119894 119895)

(6)

119883lomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1Bitmap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BitMap119909(119894 119895)) lowast 119883 (119894 119895)

(7)

lowast119909 = 119877 119866 and 119861 lowast(5) Identify the higher bit planes for each color

component starting from bit plane 5 to bit plane8 which is equal to 1 Consider

119909bp = 119909bp(bp5=1sdotsdotsdotorsdotsdotsdotbp6=1sdotsdotsdotorsdotsdotsdotbp7=1sdotsdotsdotorsdotsdotsdotbp8=1) (8)

lowast119909 = 119877 119866 and 119861 lowast(6) Compute local threshold for the identified sig-

nificant bit planes for each color componentusing Niblackrsquos method Consider

119879119909bp = ( 1

119898 lowast 119899) lowast

119898

sum

119894=1

119899

sum

119895=1

119909 (119894 119895) + 119896 lowast stddev (9)

where 119896 = 06lowast119909 = 119877 119866 and 119861 lowast

4 Journal of Engineering

(a) (b) (c)

(d) (e) (f)

Figure 2 (a) Original image (b) image of bit plane 5 (c) image of bit plane 6 (d) image of bit plane 7 (e) image of bit plane 8 and (f)amalgamated image of bit planes 5 6 7 and 8

(a) (b) (c) (d) (e)

Figure 3 (a) Original image (b) image of bit plane 5 (c) binarization of original image using mean threshold (d) binarization of bit plane 5with mean threshold and (e) binarization of bit plane 5 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 4 (a) Original image (b) image of bit plane 6 (c) binarization of original image using mean threshold (d) binarization of bit plane6 with mean threshold and (e) binarization of bit plane 6 with Niblackrsquos method of local threshold

Journal of Engineering 5

(a) (b) (c) (d) (e)

Figure 5 (a) Original image (b) image of bit plane 7 (c) binarization of original image using mean threshold (d) binarization of bit plane 7with mean threshold and (e) binarization of bit plane 7 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 6 (a) Original image (b) image of bit plane 8 (c) binarization of original image using mean threshold (d) binarization of bit plane8 with mean threshold and (e) binarization of bit plane 8 with Niblackrsquos method of local threshold

(7) Compute binary image maps for each pixel ofsignificant bit planes for 119877 119866 and 119861 respec-tively Consider

Bpbitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 119879119909bp0 if 119909 (119894 119895) lt 119879119909bp

(10)

lowast119909 = 119877 119866 and 119861 lowast(8) Compute the feature vectors 119909bitplanehi and

119909bitplanelo for bit planes for all the three colorcomponents Consider

119883bpupmean = 1

sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BPBitMap119909(119894 119895) lowast 119883 (119894 119895)

(11)

119883bplomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BPBitMap119909(119894 119895)) lowast 119883 (119894 119895)

(12)

lowast119909 = 119877 119866 and 119861 lowast(9) The feature vectors 119883upmean 119883lomean

119883bpupmean and 119883bplomean for each colorcomponent are associated to form twelvefeature vectors altogether for each image in thedataset

End

4 Experimental Process

The proposed technique was experimented with a subsetof Corel stock photo database known as Wang datasetused by Li and Wang in [18] It has been considered as awidely used public dataset The experimental process hasnot compromised with the size and quality of the imagesThe dataset comprised 9 categories with 100 images in eachcategory Figure 7 shows a sample of the original databaseThe classification performances were evaluated with 10-foldcross-validation scheme for the proposed and existing featurevector extraction techniques The process has been called 119899-fold cross-validation as given by Sridhar in [19] The valueof 119899 is an integer and was considered to be 10 in this workThe entire dataset was divided into 10 subsets 1 subset wasconsidered as the testing set and the remaining 9 subsets wereconsidered to be training sets The method was repeated for10 trials and the performance of the classifiers was evaluatedby combining the 10 results thus obtained after evaluating the10 folds

5 Evaluation of Proposed Technique

The proposed technique for feature extraction has beenevaluated by mean square error (MSE) method as in (13)The method considered the similarity measure between twoinstances as described by Xu et al in [20] and Kotsiantisin [21] The nearest neighbor in the instance space waslocated for classification and then the unknown instancewas designated with the same class of the identified nearestneighbor Consider

MSE = 1

119872119873

119872

sum

119910=1

119873

sum

119909=1

[119868 (119909 119910) minus 1198681015840(119909 119910)]

2

(13)

6 Journal of Engineering

Table 1 Comparison of misclassification rate (MR) for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses MountainsMisclassification rate (MR) 008 011 015 007 000 006 002 002 012

Table 2 Comparison of 1198651 score for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses Mountains1198651 score 064 050 038 070 100 072 091 093 042

Table 3 Confusion matrix

Cat1 Cat2 Cat3 Cat4 Cat5 Cat6 Cat7 Cat8 Cat9 Classified as61 2 15 11 1 8 0 2 0 Tribals3 51 12 3 0 8 1 0 22 Sea beach9 10 41 15 0 4 4 0 17 Gothic structure8 3 13 71 0 0 0 0 5 Bus0 0 0 0 100 0 0 0 0 Dinosaur6 9 8 0 0 72 0 2 3 Elephant2 1 3 0 0 1 89 4 0 Roses0 2 1 0 0 2 0 95 0 Horses1 26 23 4 0 5 1 1 39 Mountains

Figure 7 Sample images of different categories from Wangdatabase

where 119868 and 1198681015840 are the two images used for comparison usingthe MSE method

Primarily two different evaluation metrics namely mis-classification rate (MR) and 1198651 score were considered tocompare the proposed feature extraction technique withrespect to the existing techniques

51MisclassificationRate (MR) Theerror rate of the classifierindicates the proportion of instances that have been wronglyclassified as in

MR =FP + FN

TP + TN + FP + FN (14)

where true positive (TP) is number of instances classifiedcorrectly true negative (TN) is number of negative resultscreated for negative instances false positive (FP) is numberof erroneous results as positive results for negative instancesand false negative (FN) is number of erroneous results asnegative results for positive instances

52 F1-Score Precision andRecall (TP rate) can be combinedto produce a metric known as 1198651 score as in (15) It has beenconsidered as the harmonic mean of Precision and RecallHigher value of1198651 score indicates better classification resultsConsider

1198651score = 2 lowast Precision lowast RecallPrecision + Recall

(15)

where Precision is the probability that an object is classifiedcorrectly as per the actual value and Recall is the probabilityof a classifier that it will produce true positive result

The comparison shown in Tables 1 and 2 has beengraphically represented in Figures 8 and 9 It has beenobserved that the images of category named dinosaur havegot minimummisclassification rate and the highest 1198651 scoreOn the other hand least classification performance has beenexhibited by the category named gothic structure as shown inFigures 8 and 9The category named gothic structure has thehighest misclassification rate (MR) and the lowest 1198651 scorewhich signified poor feature extraction compared to the othercategories in the dataset The confusion matrix for all thecategories has been given as in Table 3

6 Results and Discussion

The proposed technique has implemented feature extractionby binarization of significant bit planes with Niblackrsquos local

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Journal of Engineering

(a) (b) (c)

(d) (e) (f)

Figure 2 (a) Original image (b) image of bit plane 5 (c) image of bit plane 6 (d) image of bit plane 7 (e) image of bit plane 8 and (f)amalgamated image of bit planes 5 6 7 and 8

(a) (b) (c) (d) (e)

Figure 3 (a) Original image (b) image of bit plane 5 (c) binarization of original image using mean threshold (d) binarization of bit plane 5with mean threshold and (e) binarization of bit plane 5 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 4 (a) Original image (b) image of bit plane 6 (c) binarization of original image using mean threshold (d) binarization of bit plane6 with mean threshold and (e) binarization of bit plane 6 with Niblackrsquos method of local threshold

Journal of Engineering 5

(a) (b) (c) (d) (e)

Figure 5 (a) Original image (b) image of bit plane 7 (c) binarization of original image using mean threshold (d) binarization of bit plane 7with mean threshold and (e) binarization of bit plane 7 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 6 (a) Original image (b) image of bit plane 8 (c) binarization of original image using mean threshold (d) binarization of bit plane8 with mean threshold and (e) binarization of bit plane 8 with Niblackrsquos method of local threshold

(7) Compute binary image maps for each pixel ofsignificant bit planes for 119877 119866 and 119861 respec-tively Consider

Bpbitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 119879119909bp0 if 119909 (119894 119895) lt 119879119909bp

(10)

lowast119909 = 119877 119866 and 119861 lowast(8) Compute the feature vectors 119909bitplanehi and

119909bitplanelo for bit planes for all the three colorcomponents Consider

119883bpupmean = 1

sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BPBitMap119909(119894 119895) lowast 119883 (119894 119895)

(11)

119883bplomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BPBitMap119909(119894 119895)) lowast 119883 (119894 119895)

(12)

lowast119909 = 119877 119866 and 119861 lowast(9) The feature vectors 119883upmean 119883lomean

119883bpupmean and 119883bplomean for each colorcomponent are associated to form twelvefeature vectors altogether for each image in thedataset

End

4 Experimental Process

The proposed technique was experimented with a subsetof Corel stock photo database known as Wang datasetused by Li and Wang in [18] It has been considered as awidely used public dataset The experimental process hasnot compromised with the size and quality of the imagesThe dataset comprised 9 categories with 100 images in eachcategory Figure 7 shows a sample of the original databaseThe classification performances were evaluated with 10-foldcross-validation scheme for the proposed and existing featurevector extraction techniques The process has been called 119899-fold cross-validation as given by Sridhar in [19] The valueof 119899 is an integer and was considered to be 10 in this workThe entire dataset was divided into 10 subsets 1 subset wasconsidered as the testing set and the remaining 9 subsets wereconsidered to be training sets The method was repeated for10 trials and the performance of the classifiers was evaluatedby combining the 10 results thus obtained after evaluating the10 folds

5 Evaluation of Proposed Technique

The proposed technique for feature extraction has beenevaluated by mean square error (MSE) method as in (13)The method considered the similarity measure between twoinstances as described by Xu et al in [20] and Kotsiantisin [21] The nearest neighbor in the instance space waslocated for classification and then the unknown instancewas designated with the same class of the identified nearestneighbor Consider

MSE = 1

119872119873

119872

sum

119910=1

119873

sum

119909=1

[119868 (119909 119910) minus 1198681015840(119909 119910)]

2

(13)

6 Journal of Engineering

Table 1 Comparison of misclassification rate (MR) for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses MountainsMisclassification rate (MR) 008 011 015 007 000 006 002 002 012

Table 2 Comparison of 1198651 score for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses Mountains1198651 score 064 050 038 070 100 072 091 093 042

Table 3 Confusion matrix

Cat1 Cat2 Cat3 Cat4 Cat5 Cat6 Cat7 Cat8 Cat9 Classified as61 2 15 11 1 8 0 2 0 Tribals3 51 12 3 0 8 1 0 22 Sea beach9 10 41 15 0 4 4 0 17 Gothic structure8 3 13 71 0 0 0 0 5 Bus0 0 0 0 100 0 0 0 0 Dinosaur6 9 8 0 0 72 0 2 3 Elephant2 1 3 0 0 1 89 4 0 Roses0 2 1 0 0 2 0 95 0 Horses1 26 23 4 0 5 1 1 39 Mountains

Figure 7 Sample images of different categories from Wangdatabase

where 119868 and 1198681015840 are the two images used for comparison usingthe MSE method

Primarily two different evaluation metrics namely mis-classification rate (MR) and 1198651 score were considered tocompare the proposed feature extraction technique withrespect to the existing techniques

51MisclassificationRate (MR) Theerror rate of the classifierindicates the proportion of instances that have been wronglyclassified as in

MR =FP + FN

TP + TN + FP + FN (14)

where true positive (TP) is number of instances classifiedcorrectly true negative (TN) is number of negative resultscreated for negative instances false positive (FP) is numberof erroneous results as positive results for negative instancesand false negative (FN) is number of erroneous results asnegative results for positive instances

52 F1-Score Precision andRecall (TP rate) can be combinedto produce a metric known as 1198651 score as in (15) It has beenconsidered as the harmonic mean of Precision and RecallHigher value of1198651 score indicates better classification resultsConsider

1198651score = 2 lowast Precision lowast RecallPrecision + Recall

(15)

where Precision is the probability that an object is classifiedcorrectly as per the actual value and Recall is the probabilityof a classifier that it will produce true positive result

The comparison shown in Tables 1 and 2 has beengraphically represented in Figures 8 and 9 It has beenobserved that the images of category named dinosaur havegot minimummisclassification rate and the highest 1198651 scoreOn the other hand least classification performance has beenexhibited by the category named gothic structure as shown inFigures 8 and 9The category named gothic structure has thehighest misclassification rate (MR) and the lowest 1198651 scorewhich signified poor feature extraction compared to the othercategories in the dataset The confusion matrix for all thecategories has been given as in Table 3

6 Results and Discussion

The proposed technique has implemented feature extractionby binarization of significant bit planes with Niblackrsquos local

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Engineering 5

(a) (b) (c) (d) (e)

Figure 5 (a) Original image (b) image of bit plane 7 (c) binarization of original image using mean threshold (d) binarization of bit plane 7with mean threshold and (e) binarization of bit plane 7 with Niblackrsquos method of local threshold

(a) (b) (c) (d) (e)

Figure 6 (a) Original image (b) image of bit plane 8 (c) binarization of original image using mean threshold (d) binarization of bit plane8 with mean threshold and (e) binarization of bit plane 8 with Niblackrsquos method of local threshold

(7) Compute binary image maps for each pixel ofsignificant bit planes for 119877 119866 and 119861 respec-tively Consider

Bpbitmap119909(119894 119895) =

1 if 119909 (119894 119895) gt 119879119909bp0 if 119909 (119894 119895) lt 119879119909bp

(10)

lowast119909 = 119877 119866 and 119861 lowast(8) Compute the feature vectors 119909bitplanehi and

119909bitplanelo for bit planes for all the three colorcomponents Consider

119883bpupmean = 1

sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

BPBitMap119909(119894 119895) lowast 119883 (119894 119895)

(11)

119883bplomean = 1

119898 lowast 119899 minus sum119898

119894=1sum119899

119895=1BPBitMap

119909(119894 119895)

lowast

119898

sum

119894=1

119899

sum

119895=1

(1 minus BPBitMap119909(119894 119895)) lowast 119883 (119894 119895)

(12)

lowast119909 = 119877 119866 and 119861 lowast(9) The feature vectors 119883upmean 119883lomean

119883bpupmean and 119883bplomean for each colorcomponent are associated to form twelvefeature vectors altogether for each image in thedataset

End

4 Experimental Process

The proposed technique was experimented with a subsetof Corel stock photo database known as Wang datasetused by Li and Wang in [18] It has been considered as awidely used public dataset The experimental process hasnot compromised with the size and quality of the imagesThe dataset comprised 9 categories with 100 images in eachcategory Figure 7 shows a sample of the original databaseThe classification performances were evaluated with 10-foldcross-validation scheme for the proposed and existing featurevector extraction techniques The process has been called 119899-fold cross-validation as given by Sridhar in [19] The valueof 119899 is an integer and was considered to be 10 in this workThe entire dataset was divided into 10 subsets 1 subset wasconsidered as the testing set and the remaining 9 subsets wereconsidered to be training sets The method was repeated for10 trials and the performance of the classifiers was evaluatedby combining the 10 results thus obtained after evaluating the10 folds

5 Evaluation of Proposed Technique

The proposed technique for feature extraction has beenevaluated by mean square error (MSE) method as in (13)The method considered the similarity measure between twoinstances as described by Xu et al in [20] and Kotsiantisin [21] The nearest neighbor in the instance space waslocated for classification and then the unknown instancewas designated with the same class of the identified nearestneighbor Consider

MSE = 1

119872119873

119872

sum

119910=1

119873

sum

119909=1

[119868 (119909 119910) minus 1198681015840(119909 119910)]

2

(13)

6 Journal of Engineering

Table 1 Comparison of misclassification rate (MR) for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses MountainsMisclassification rate (MR) 008 011 015 007 000 006 002 002 012

Table 2 Comparison of 1198651 score for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses Mountains1198651 score 064 050 038 070 100 072 091 093 042

Table 3 Confusion matrix

Cat1 Cat2 Cat3 Cat4 Cat5 Cat6 Cat7 Cat8 Cat9 Classified as61 2 15 11 1 8 0 2 0 Tribals3 51 12 3 0 8 1 0 22 Sea beach9 10 41 15 0 4 4 0 17 Gothic structure8 3 13 71 0 0 0 0 5 Bus0 0 0 0 100 0 0 0 0 Dinosaur6 9 8 0 0 72 0 2 3 Elephant2 1 3 0 0 1 89 4 0 Roses0 2 1 0 0 2 0 95 0 Horses1 26 23 4 0 5 1 1 39 Mountains

Figure 7 Sample images of different categories from Wangdatabase

where 119868 and 1198681015840 are the two images used for comparison usingthe MSE method

Primarily two different evaluation metrics namely mis-classification rate (MR) and 1198651 score were considered tocompare the proposed feature extraction technique withrespect to the existing techniques

51MisclassificationRate (MR) Theerror rate of the classifierindicates the proportion of instances that have been wronglyclassified as in

MR =FP + FN

TP + TN + FP + FN (14)

where true positive (TP) is number of instances classifiedcorrectly true negative (TN) is number of negative resultscreated for negative instances false positive (FP) is numberof erroneous results as positive results for negative instancesand false negative (FN) is number of erroneous results asnegative results for positive instances

52 F1-Score Precision andRecall (TP rate) can be combinedto produce a metric known as 1198651 score as in (15) It has beenconsidered as the harmonic mean of Precision and RecallHigher value of1198651 score indicates better classification resultsConsider

1198651score = 2 lowast Precision lowast RecallPrecision + Recall

(15)

where Precision is the probability that an object is classifiedcorrectly as per the actual value and Recall is the probabilityof a classifier that it will produce true positive result

The comparison shown in Tables 1 and 2 has beengraphically represented in Figures 8 and 9 It has beenobserved that the images of category named dinosaur havegot minimummisclassification rate and the highest 1198651 scoreOn the other hand least classification performance has beenexhibited by the category named gothic structure as shown inFigures 8 and 9The category named gothic structure has thehighest misclassification rate (MR) and the lowest 1198651 scorewhich signified poor feature extraction compared to the othercategories in the dataset The confusion matrix for all thecategories has been given as in Table 3

6 Results and Discussion

The proposed technique has implemented feature extractionby binarization of significant bit planes with Niblackrsquos local

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Journal of Engineering

Table 1 Comparison of misclassification rate (MR) for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses MountainsMisclassification rate (MR) 008 011 015 007 000 006 002 002 012

Table 2 Comparison of 1198651 score for the nine image categories

Categories Tribals Sea beach Gothic structure Bus Dinosaur Elephant Roses Horses Mountains1198651 score 064 050 038 070 100 072 091 093 042

Table 3 Confusion matrix

Cat1 Cat2 Cat3 Cat4 Cat5 Cat6 Cat7 Cat8 Cat9 Classified as61 2 15 11 1 8 0 2 0 Tribals3 51 12 3 0 8 1 0 22 Sea beach9 10 41 15 0 4 4 0 17 Gothic structure8 3 13 71 0 0 0 0 5 Bus0 0 0 0 100 0 0 0 0 Dinosaur6 9 8 0 0 72 0 2 3 Elephant2 1 3 0 0 1 89 4 0 Roses0 2 1 0 0 2 0 95 0 Horses1 26 23 4 0 5 1 1 39 Mountains

Figure 7 Sample images of different categories from Wangdatabase

where 119868 and 1198681015840 are the two images used for comparison usingthe MSE method

Primarily two different evaluation metrics namely mis-classification rate (MR) and 1198651 score were considered tocompare the proposed feature extraction technique withrespect to the existing techniques

51MisclassificationRate (MR) Theerror rate of the classifierindicates the proportion of instances that have been wronglyclassified as in

MR =FP + FN

TP + TN + FP + FN (14)

where true positive (TP) is number of instances classifiedcorrectly true negative (TN) is number of negative resultscreated for negative instances false positive (FP) is numberof erroneous results as positive results for negative instancesand false negative (FN) is number of erroneous results asnegative results for positive instances

52 F1-Score Precision andRecall (TP rate) can be combinedto produce a metric known as 1198651 score as in (15) It has beenconsidered as the harmonic mean of Precision and RecallHigher value of1198651 score indicates better classification resultsConsider

1198651score = 2 lowast Precision lowast RecallPrecision + Recall

(15)

where Precision is the probability that an object is classifiedcorrectly as per the actual value and Recall is the probabilityof a classifier that it will produce true positive result

The comparison shown in Tables 1 and 2 has beengraphically represented in Figures 8 and 9 It has beenobserved that the images of category named dinosaur havegot minimummisclassification rate and the highest 1198651 scoreOn the other hand least classification performance has beenexhibited by the category named gothic structure as shown inFigures 8 and 9The category named gothic structure has thehighest misclassification rate (MR) and the lowest 1198651 scorewhich signified poor feature extraction compared to the othercategories in the dataset The confusion matrix for all thecategories has been given as in Table 3

6 Results and Discussion

The proposed technique has implemented feature extractionby binarization of significant bit planes with Niblackrsquos local

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Engineering 7

016

014

012

01

008

006

004

002

0

Misc

lass

ifica

tion

rate

(MR)

Misclassification rate (MR)

Tribals Sea beach Bus Dinosaur Elephant Roses Horses Mountains

008 011 015 007 0 006 002 002 012

Gothicstructure

Category-wise misclassification rate comparison with proposed technique

Figure 8 Comparison of misclassification rate (MR) for different categories

0

02

04

06

08

1

12

064 05 038 07 1 072 091 093 042

scor

e

score

Tribals Sea beach Bus Dinosaur Elephant Roses Horses MountainsGothicstructure

F1

F1

Category-wiseF1 score comparison with proposed technique

Figure 9 Comparison of 1198651 score for different categories

threshold selection method Binarization of significant bitplanes was done in [2] withmean thresholdmethod Anotherexisting technique of feature extraction in [3] has also usedmean threshold for binarization Comparative analysis ofmisclassification rate (MR) and 1198651 score of the proposedtechnique of feature extraction has outperformed the featureextraction method of [2 3] as shown in Table 4 and Figures10 and 11

The proposed binarization technique of feature extrac-tion has also been compared with widely used global andlocal threshold techniques for binarization namely Otsursquosmethod Niblackrsquos method Sauvolarsquos method and Bernsenrsquosmethod The results in Table 4 and Figures 10 and 11 have

revealed minimum misclassification rate (MR) and maxi-mum 1198651 score respectively for the proposed technique withrespect to all the techniques compared Table 4 has shownthe comparison of proposed technique with reference to theexisting techniques in terms of misclassification rate and1198651 score Minimum misclassification rate of 007 for thetest image categories has been observed with the proposedmethod The existing method of feature extraction by bina-rization of significant bit planes [2] using mean threshold hasshown higher misclassification rate (MR) of 0078 comparedto the proposed technique The 1198651 score of the proposedmethod was the highest among all the existing techniquesThe proposed technique for feature extraction was further

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Journal of Engineering

Table4Com

paris

onof

misc

lassificatio

nrate(M

R)and1198651score

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Misc

lassificatio

nrate(M

R)with

KNNcla

ssifier

007

0075

0078

0078

008

0081

0096

0107

1198651score

with

KNNcla

ssifier

069

066

065

065

064

063

057

052

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Engineering 9

012

01

008

006

004

002

0

Techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Misc

lass

ifica

tion

rate

(MR)

Comparison of misclassification rate (MR) of the proposedtechnique and the existing techniques

Figure 10 Comparison of misclassification rate (MR) of theproposed technique with respect to the existing techniques

compared to the existing state-of-the-art techniques withrespect to average time taken for extraction of features fromeach image in Wang database The comparison has beenshown in Figure 12 It was observed that the proposed tech-nique has consumed minimum time for feature extractioncompared to the existing techniques

Feature extraction by binarization with multilevel meanthreshold has taken the maximum time followed by Sauvolarsquoslocal threshold method for feature extraction Subsequenttime was consumed by feature extraction with Bernsenrsquoslocal threshold method Niblackrsquos local threshold methodand Otsursquos global threshold method respectively The twotechniques namely feature extraction with Bit Plane Slicingusing mean threshold and feature extraction by binarizationof original + even image with mean threshold respectively

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction with binarization withBernsenrsquos local threshold method (existing)

Feature extraction with binarization withSauvolarsquos local threshold method (existing)Feature extraction with binarization withNiblackrsquos local threshold method (existing)Feature extraction with binarization withOtsursquos global threshold method (existing)

Techniques

08

06

04

02

01

03

05

07

0

scor

e

Comparison of score of the proposedtechnique and the existing techniques

F1

F1

Figure 11 Comparison of 1198651 score of the proposed technique withrespect to the existing techniques

have lesser time consumption compared to the rest of theexisting techniques

The results clearly established better classification perfor-mance of the proposed technique with respect to the existingmean threshold based techniques and the traditional globaland local threshold techniques adopted for binarization tofacilitate feature extraction

Table 5 has shown the comparison of Precision Recalland Accuracy of the proposed technique with respect to theexisting methods of binarization for feature extraction

The graphical comparison shown in Figure 13 has clearlyascertained that the proposed technique has highest amountof precision and recall value with maximum accuracy amongall the techniques compared

7 Analysis of Statistical Comparison

The output from various evaluation metrics was individuallyintegrated to assess the relationship of each metric used

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

10 Journal of Engineering

Table5Com

paris

onof

Precision

RecallandAc

curacy

ofdifferent

techniqu

es

Techniqu

esProp

osed

Featuree

xtraction

bybinariz

ation

with

multilevel

meanthreshold

(existing

)

Featuree

xtraction

bybinariz

ation

usingBitP

lane

Slicingwith

mean

threshold

(existing

)

Featuree

xtraction

bybinariz

ationof

original+even

imagew

ithmean

threshold

(existing

)

Featuree

xtraction

bybinariz

ation

with

Bernsenrsquos

localthresho

ldmetho

d(existing)

Featuree

xtraction

bybinariz

ation

with

Sauvolarsquos

local

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Niblackrsquoslocal

thresholdmetho

d(existing

)

Featuree

xtraction

bybinariz

ation

with

Otsu

rsquosglob

althresholdmetho

d(existing

)

Precision

069

066

065

065

064

063

057

052

Recall

069

066

065

065

064

063

057

052

Accuracy

093

092

092

092

092

092

09

089

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Engineering 11

Table 6 Test of correlation among different evaluation metrics

Test of correlationsMisclassification rate (MR)

with KNN classifier1198651 score withKNN classifier Precision Recall Accuracy

Misclassification rate (MR)with KNN classifier Pearson correlation 1

1198651 score with KNNclassifier Pearson correlation minus998lowastlowast 1

Precision Pearson correlation minus998lowastlowast 1000lowastlowast 1Recall Pearson correlation minus998lowastlowast 1000lowastlowast 1000lowastlowast 1Accuracy Pearson correlation minus989lowastlowast 986lowastlowast 986lowastlowast 986lowastlowast 1lowastlowastCorrelation is significant at the 001 level (2-tailed)

Comparison of average rate computation time required for feature extraction from each image

Tim

e (s)

Average computation time (s)

0

2

4

6

8

10

12

14

Proposed

Featureextraction

bybinarization

withmultilevel

meanthreshold(existing)

Featureextraction

bybinarization

usingbit planeslicingwithmean

threshold(existing)

Featureextraction

bybinarization

withBernsenrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withSauvolarsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withNiblackrsquos

localthresholdmethod

(existing)

Featureextraction

bybinarization

withOtsursquoslocal

thresholdmethod

(existing)

Featureextraction

bybinarization

of

imagewithmean

threshold(existing)

original +even

1593 13025 2040 1934 8516 8811 7961 3906

Figure 12 Comparison of averge computation time for feature extraction from each image

for performance measure Correlation analysis was usedto explore the relationship among various metrics namelymisclassification rate (MR) 1198651 score Precision Recall andAccuracy as suggested by Bishara and Hittner [22] Table 6has shown that 1198651 score Precision Recall and Accuracyare negatively correlated with misclassification rate (MR)Precision Recall and Accuracy are highly correlated with1198651 score Thus it was inferred that any method must try tominimize the misclassification rate for increased 1198651 scorePrecision Recall and Accuracy

The proposed feature extraction technique was comparedwith seven existing techniques of feature extraction for classi-fication performance evaluationThe correlation analysis hasclearly established the necessity for minimized misclassifica-tion rate for better classification The authors have comparedthe feature extraction techniques in terms of false positiveand false negative results generated during classification foreach of the nine categories considered in Wangrsquos dataset

Table 7 2 times 2 confusion matrix

True class Predicted class SumPositive Negative

Positive TP FN pNegative FP TN nSum 119901

10158401198991015840

Aunivariate 119905 test (one-tailed) was conducted for comparisonof each of the existing feature extraction techniques to theproposed technique in terms of pair as proposed by Yıldızet al [23] and Sharma [24] Two algorithms in the pair weretrained and validated on 119896 training data folds and 2 times 2

confusion matrices 119872119894119895 119894 = 1 2 and 119895 = 1 119896 were

calculated as shown in Table 7Comparison was done in terms of errors and 119890

119894119895= fp119894119895+

fn119894119895was calculated for each of the algorithms It was followed

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

12 Journal of Engineering

Table 8 Univariate 119905 test

Comparison 119905-calc 119875 value SignificanceFeature extraction by binarization with multilevel mean threshold (existing) minus261320 0015488 lowast

Feature extraction by binarization using Bit Plane Slicing with mean threshold (existing) minus250732 0018261 lowast

Feature extraction by binarization of original + even image with mean threshold (existing) minus262862 0015121 lowast

Feature extraction by binarization with Bernsenrsquos local threshold method (existing) minus266076 0014386 lowast

Feature extraction by binarization with Sauvolarsquos local threshold method (existing) minus241902 0020957 lowast

Feature extraction by binarization with Niblackrsquos local threshold method (existing) minus369717 0003034 lowastlowast

Feature extraction by binarization with Otsursquos global threshold method (existing) minus520553 0000408 lowastlowast

lowastStands for significant and lowastlowastfor highly significant

1

09

08

07

06

05

04

03

02

01

0

RecallPrecision Accuracy

technique and the existing techniques

Proposed

Feature extraction by binarization withmultilevel mean threshold (existing)Feature extraction by binarization usingbit plane slicing with mean threshold (existing)

Feature extraction by binarization oforiginal + even image with mean threshold (existing)Feature extraction by binarization withBernsenrsquos local threshold method (existing)

Feature extraction by binarization withSauvolarsquos local threshold method (existing)Feature extraction by binarization withNiblackrsquos local threshold method (existing)Feature extraction by binarization withOtsursquos local threshold method (existing)

Comparison of Precision Recall and Accuracy of the proposed

Figure 13 Comparison of Precision Recall and Accuracy of theproposed technique with respect to the existing techniques

by calculation of paired difference between the errors 119889119895=

1198901119895minus 1198902119895 The test calculated whether the differences were

generated from a population with zero mean

1198670 120583119889 = 0 versus 1198671 120583119889 lt 0 (16)

The results for comparison for each of the techniques to theproposed technique have been shown in Table 8

Analyses shown in Table 8 have indicated the 119875 val-ues as significant or highly significant Therefore the nullhypotheses of equal error rates for the existing algorithmcompared to the proposed algorithm were rejected As suchwe can conclude that the error rate of the proposed techniquein comparison to existing techniques is significantly lessHence the proposed technique has contributed considerableimprovement in classification performance compared to theexisting techniques of feature extraction

8 Conclusion

Thepaper has presented a new technique of feature extractionwith the help of image binarization The statistical com-parison has shown significant improvement in classifica-tion performance for the proposed technique compared tothe existing feature extraction techniques by binarizationwith mean threshold method global threshold method andlocal threshold method The work has the possibility to beextended for a large variety of applications including citysurveillance and medical image analysis where binarizationis very important as a preprocessing step for consequentobject recognition The proposed method has demonstratedbetter average results for five different performance evalu-ation measures compared to the other widely used featureextraction methods Thus the novel technique of featureextraction with significant bit planes using binarization withlocal threshold has been established as a consistent techniquefor feature extraction in content based image classification

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] A Andreopoulos and J K Tsotsos ldquo50 Years of objectrecognition directions forwardrdquo Computer Vision and ImageUnderstanding vol 117 no 8 pp 827ndash891 2013

[2] S Thepade R Das and S Ghosh ldquoPerformance comparison offeature vector extraction techniques in RGB color space usingblock truncation coding for content based image classificationwith discrete classifiersrdquo inProceedings of the Annual IEEE IndiaConference (INDICON rsquo13) pp 1ndash6 December 2013

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Engineering 13

[3] H B Kekre S Thepade R K Das and S Ghosh ldquoMultilevelblock truncation coding with diverse color spaces for imageclassificationrdquo in Proceedings of the International Conference onAdvances in Technology and Engineering (ICATE rsquo13) pp 1ndash7January 2013

[4] WNiblackAn Introduction toDigital Image Processing PrenticeHall Englewood Cliffs NJ USA 1986

[5] J Bernsen ldquoDynamic thresholding of gray level imagesrdquo in Pro-ceedings of the International Conference on Pattern Recognition(ICPR rsquo86) pp 1251ndash1255 1986

[6] J Sauvola and M Pietikainen ldquoAdaptive document imagebinarizationrdquo Pattern Recognition vol 33 no 2 pp 225ndash2362000

[7] N Otsu ldquoA threshold selection method from gray-level his-togramrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

[8] M Valizadeh N Armanfard M Komeili and E Kabir ldquoAnovel hybrid algorithm for binarization of badly illuminateddocument imagesrdquo in Proceedings of the 14th International CSIComputer Conference (CSICC rsquo09) pp 121ndash126 Tehran IranOctober 2009

[9] Y-F Chang Y-T Pai and S-J Ruan ldquoAn efficient thresholdingalgorithm for degraded document images based on intelligentblock detectionrdquo in Proceedings of the IEEE International Con-ference on Systems Man and Cybernetics (SMC rsquo08) pp 667ndash672 October 2009

[10] B Gatos I Pratikakis and S J Perantonis ldquoEfficient bina-rization of historical and degraded document imagesrdquo in Pro-ceedings of the 8th IAPR International Workshop on DocumentAnalysis Systems (DAS rsquo08) pp 447ndash454 September 2008

[11] M E Elalami ldquoA novel image retrievalmodel based on themostrelevant featuresrdquo Knowledge-Based Systems vol 24 no 1 pp23ndash32 2011

[12] P S Hiremath and J Pujari ldquoContent based image retrievalusing color texture and shape featuresrdquo in Proceedings of the15th International Conference on Advanced Computing andCommunication (ADCOM rsquo07) pp 780ndash784 December 2007

[13] M Banerjee M K Kundu and P Maji ldquoContent-based imageretrieval using visually significant point featuresrdquoFuzzy Sets andSystems vol 160 no 23 pp 3323ndash3341 2009

[14] H A Jalab ldquoImage retrieval system based on color layoutdescriptor and Gabor filtersrdquo in Proceedings of the IEEE Con-ference on Open Systems (ICOS rsquo11) pp 32ndash36 IEEE LangkawiMalaysia September 2011

[15] G-L Shen and X-J Wu ldquoContent based image retrieval bycombining color texture and CENTRISTrdquo in Proceedings ofthe Constantinides International Workshop on Signal Processing(CIWSP rsquo13) pp 1ndash4 January 2013

[16] A Irtaza M A Jaffar E Aleisa and T S Choi ldquoEmbeddingneural networks for semantic association in content basedimage retrievalrdquo Multimedia Tools and Applications pp 1ndash212013

[17] M Rahimi and M E Moghaddam ldquoA content-based imageretrieval system based on color ton distribution descriptorsrdquoSignal Image and Video Processing 2013

[18] J Li and J ZWang ldquoAutomatic linguistic indexing of pictures bya statistical modeling approachrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 25 no 9 pp 1075ndash10882003

[19] S Sridhar ldquoImage features representation and descriptionrdquo inDigital Image Processing pp 483ndash486 India Oxford UniversityPress New Delhi India 2011

[20] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[21] S B Kotsiantis ldquoSupervised machine learning a review ofclassification techniquesrdquo Informatica vol 31 no 3 pp 249ndash268 2007

[22] A J Bishara and J B Hittner ldquoTesting the significance ofa correlation with nonnormal data comparison of PearsonSpearman transformation and resampling approachesrdquo Psy-chological Methods vol 17 no 3 pp 399ndash417 2012

[23] O T Yıldız O Aslan and E Alpaydın ldquoMultivariate statisticaltests for comparing classification algorithmsrdquo in Learning andIntelligent Optimization vol 6683 of Lecture Notes in ComputerScience pp 1ndash15 Springer Berlin Germany 2011

[24] J K Sharma Fundamentals of Business Statistics Vikash Pub-lishing House 2nd edition 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of