research article text extraction from historical document

Research ArticleText Extraction from Historical Document Images bythe Combination of Several Thresholding Techniques

Toufik Sari1 Abderrahmane Kefali12 and Halima Bahi1

1 LabGED Laboratory Badji Mokhtar BP 12 23000 Annaba Algeria2 08 May 1945 University BP 401 24000 Guelma Algeria

Correspondence should be addressed to Abderrahmane Kefali kefalilabgednet

Received 13 June 2014 Revised 4 September 2014 Accepted 10 September 2014 Published 29 September 2014

Academic Editor Chengcui Zhang

Copyright copy 2014 Toufik Sari et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

This paper presents a new technique for the binarization of historical document images characterized by deteriorations and damagesmaking their automatic processing difficult at several levels The proposed method is based on hybrid thresholding combining theadvantages of global and local methods and on the mixture of several binarization techniques Two stages have been included Inthe first stage global thresholding is applied on the entire image and two different thresholds are determined from which the mostof image pixels are classified into foreground or background In the second stage the remaining pixels are assigned to foregroundor background classes based on local analysis In this stage several local thresholding methods are combined and the final binaryvalue of each remaining pixel is chosen as the most probable one The proposed technique has been tested on a large collection ofstandard and synthetic documents and compared with well-known methods using standard measures and was shown to be morepowerful

1 Introduction

Binarization is an important step in the process of documentanalysis and recognition It has as goal to segment the imageinto two classes (foreground and background in the case ofdocument images) The resulting image is a binary image inblack and white where the black represents the foregroundand the white represents the background In fact documentimage binarization is critical in the sense that bad separationwill cause the loss of pertinent information andor adduseless information (noise) generating wrong results Thisdifficulty increases for old documents which have varioustypes of damages and degradations from the digitizationprocess itself aging effects humidity marks fungus dirt andso forth making the automatic processing of these materialsdifficult at several levels

A great number of techniques have been proposed inthe literature for the binarization of gray-scale or coloreddocuments images but no one between them is generic andefficient for all types of documents

The binarization techniques of grayscale images may beclassified into two categories global thresholding and local

thresholding [1 2] Another category of hybrid methodscan be added [3] The global thresholding methods arewidely used in many document image analysis applicationsfor their simplicity and efficiency However these methodsare powerful only when the original documents are ofgood quality and well contrasted and have a clear bimodalpattern that separates foreground text and background Forhistorical document images which are generally noisy andof poor quality global thresholding methods become notsuitable because no single threshold is able to completelyseparate the foreground from the background of the imagesince there is no sufficient distinction between the grayrange of background and foreground pixels This kind ofdocument requires a more detailed analysis which may beguaranteed by local methods Local methods calculate adifferent threshold for each pixel based on the informationof its neighborhoodsThese methods are more robust againstuneven illumination low contrast and varying colors thanglobal ones but they are very time consuming since a separatethreshold is computed for each pixel of the image consideringits neighborhoods This calculation becomes slower whenincreasing the size of neighborhood considered Hybrid

Hindawi Publishing CorporationAdvances in MultimediaVolume 2014 Article ID 934656 10 pageshttpdxdoiorg1011552014934656

2 Advances in Multimedia

methods in contrast combine global and local informationfor segmenting the image

In this paper we propose a new hybrid thresholdingtechnique for binarizing images of historical documentsThe proposed approach uses a mixture of thresholdingmethods and it combines the advantages of the two familiesof techniques the computation speed and the efficiencyThe remainder of this paper is organized as follows InSection 2 we present some existing binarization methodsThen in Section 3 we describe the proposed approach Theexperiments performed and the results will be shown inSection 4 before concluding

2 State of the Art

According to [4] existing methodologies for image bina-rization may be divided under two main strategies groupingbased and thresholding based Thresholding based methodsuse global or local threshold(s) to separate the text from thebackground In grouping based methods we distinguish twocategories region based grouping and clustering based group-ing methods Region based grouping methods are mainlybased on spatial-domain region growing or on splitting andmerging Also clustering based grouping methods are basedon classification of intensity or color values as a functionof a homogeneity criterion However several techniqueshave been employed to reach this classification K-meansalgorithm artificial neural networks and so forth

Sezgin and Sankur [5] established a classification ofbinarization methods according to the information that theyexploit in 6 categories

(i) Histogram-based methods the methods of this classperform a thresholding based on the form of thehistogram

(ii) Clustering-based methods these methods assign theimage pixels to one of the two clusters object andbackground

(iii) Entropy-based methods these algorithms use theinformation theory to obtain the threshold

(iv) Object attribute-based methods they find a thresh-old value based on some similarity measurementsbetween original and binary images

(v) Spatial binarization methods they find the optimalthreshold value taking into account spatial measures

(vi) Locally adaptive methods these methods aredesigned to give a new threshold for every pixelSeveral kinds of adaptive methods exist We findmethods based on local gray range local variationand so forth

Wepresent in this section some binarizationmethods themost frequently cited in the literature and we consider onlythresholding based methods

21 Global Methods Note 119868 the grayscale image of whichthe intensities vary from 0 (black) to 1 (white) and 119867 its

histogram of intensities The number of pixels having a graylevel 119894 is denoted as119867(119894)

211 Otsursquos Method Otsursquos method [6] tries to find thethreshold 119879 which separates the gray-level histogram inan optimal way into two segments (which maximize theintersegments variance or whichminimize the intrasegmentsvariance) The calculation of the interclasses or intraclassesvariances is based on the normalized histogram 119867

119899=

[119867119899(0) sdot sdot sdot 119867

119899(255)] of the image where sum119867

119899(119894) = 1

The interclasses variance for each gray level 119905 is given by

119881inter = 1199021(119905) times 119902

2(119905) times [120583

1(119905) minus 120583

2(119905)]2 (1)

such that

1205831(119905) =

1

1199021(119905)

119905minus1

sum

119894=0

119867119899(119894) times 119894

1205832(119905) =

1

1199022(119905)

255

sum

119894=119905

119867119899(119894) times 119894

1199021(119905) =

119905minus1

sum

119894=0

119867119899(119894) 119902

2(119905) =

255

sum

119894=119905

119867119899(119894)

(2)

212 ISODATA Method Thresholding using ISODATA [7]consists in finding a threshold by separating iteratively thegray-level histogram into two classes with the a prioriknowledge of the values associated with each class Thismethod starts by dividing the interval of nonnull values ofthe histogram into two equidistant parts and next we take119898

1

and 1198982as the arithmetic average of each class Repeat until

convergence the calculation of the optimal threshold 119879 as theclosest integer to (119898

1+1198982)2 and update the two averages119898

1

and1198982

213 Kapur et alrsquos Method Kapur et alrsquos method [8] isan entropy based method which takes into account theforeground likelihood distribution 119875

119891and the background

likelihood distribution (119875119887= 1 minus 119875

119891) in the determination of

the division entropy The binarization threshold 119879 is chosenfor which the value 119864 = 119864

119891+ 119864119887is maximal such that

119864119891= minus

119905

sum

119894=0

119875119894

119875119891

times log(119875119894

119875119891

)

119864119887= minus

255

sum

119894=119905+1

119875119894

1 minus 119875119891

times log(119875119894

1 minus 119875119891

)

(3)

where 119875119894is the occurrence probability of the gray level 119894 in the

image and 119875119891= sum119905

119894=0119875119894

214 Iterative Global Thresholding (IGT) The proposedmethod selects a global threshold to the entire image basedon an iterative procedure [9] At each iteration 119894 the followingsteps are performed

(a) calculating the average gray level (119879119894) of the image

(b) subtracting 119879119894from all pixels of the image

Advances in Multimedia 3

(c) histogram equalization to extend the pixels over thewhole gray levels interval

The algorithm stops when |119879119894minus 119879119894minus1| lt 0001

22 LocalMethods Localmethods compute a local thresholdfor each pixel by sliding a square or rectangular window overthe entire image

221 Bernsenrsquos Method It is an adaptive local method [10]Thus for each pixel of coordinates (119909 119910) the threshold isgiven by

119879 (119909 119910) =

119885low + 119885high

2

(4)

such that 119885low and 119885high are the lowest and the highest graylevels respectively in a squared window119908times119908 centered overthe pixel (119909 119910)

However if the local contrast 119862(119909 119910) = (119885high minus 119885low) isbelow a threshold 119897 (119897 = 15) then the neighborhood consistsof a single class foreground or background

222 Niblackrsquos Method The local threshold 119879(119909 119910) is calcu-lated using the mean119898 and standard deviation 120590 of all pixelsin the window (neighborhood of the pixel in question) [11]Thus the threshold 119879(119909 119910) is given by

119879 (119909 119910) = 119898 + 119896 times 120590 (5)

such that 119896 is a parameter used for determining the numberof edge pixels considered as object pixels and takes a negativevalues (119896 is fixed minus02 by authors)

223 Sauvola and Pietikainenrsquos Method Sauvola andPietikainenrsquos algorithm [3] is a modification of that ofNiblack in order to give more performance in the documentswith a background containing a light texture or too variationand uneven illumination In the modification of Sauvola thelocal binarization threshold is given by

119879 (119909 119910) = 119898 times (1 minus 119896 times (1 minus120590

119877)) (6)

where 119877 is the dynamic range of the standard deviation 120590 andthe parameter 119896 takes positives values in the interval [02 05]

224 Nick Method This method improves considerably thebinarization of lighted images and low contrasted images bydownwards moving the threshold of binarization [2] Thethreshold calculation is done as follows

119879 (119909 119910) = 119898 + 119896radic(sum1199012

119894minus 1198982

)

NP(7)

such that 119896 is the Niblack factor and varies between minus01 andminus02 according to the application need119898 is the average graylevel 119901

119894is the gray level of pixel 119894 and NP is the total number

of pixels In their tests the authors used a window of size 19times19

225 Sari et alrsquos Method This method uses an artificialneuron network of multilayer perceptron (MLP) type toclassify the image pixels into two classes foreground andbackground [12] The MLP has one hidden layer 25 inputsand one single output To assign a new value (black orwhite) to a pixel the MLP takes as input a vector of 25values corresponding to the intensities of the pixel in a5 times 5 window centered on the processed pixel The MLPparameters (structure the input statistics etc) have beenchosen after several experiments

23 Hybrid Methods

231 Improved IGTMethod Thismethod [13] is an improve-ment of IGT technique from [9] and it consists of two passesIn the first pass a global thresholding is applied to the entireimage and in the second pass a local thresholding processesareas still containing noise To do this the binary imageresulting from global thresholding is divided into severalsegments of size 119899 times 119899 and for each segment the frequency119891 of black pixels is calculated The segments satisfying thefollowing criteria are kept 119891 gt 120583 + 119896 times 120590 such that 120583 and120590 denote the mean and the standard deviation of the blackpixels frequency in the segment and 119896 is a constant (equal to2 according to the authors) For each detected area the IGTmethod is applied to the corresponding area in the originalimage Areas of size 50times50 give good results according to theauthors

232 Gangamma and Srikantarsquos Method Gangamma andSrikanta [14] proposed amethod based on a simple and effec-tive combination of spatial filters with gray scale morpho-logical operations to remove the background and improvethe quality of historical document image of palm scripts Thefirst step of this technique is to apply adaptive histogramequalization (AHE) to overcome the problem of unevenillumination in the document image On the resulting imageamorphological opening operation is applied and the openedimage is added later with the histogram equalized imageAfter that the morphological closing operation is applied tothe image for smoothing The histogram equalized image issubtracted from the smoothed image and the result is sub-tracted again from the previous addition image A Gaussianfilter is subsequently applied in order to remove the noise Afinal improvement is obtained by adding the last image withthe histogram equalized image Finally a global thresholding(Otsursquos algorithm) is required to separate the text from thebackground

233 Thresholding by Background Subtraction This tech-nique has been proposed in [15] and it consists of three stepsThe background is modeled by removing the handwritingby applying a closing of the original image with a smalldisk as a structuring element After that the background issubtracted from the original image which only leaves theforeground Finally the resulting image is segmented usingOtsursquos algorithm multiplied by an empirical constant


(a) Gray level image

Num

ber o

f pi

xels

Gray levels

133

(b) Its histogram

(c) Thresholding result with Otsursquos method

Figure 1 Global thresholding of a bimodal image

234 Tabatabaie and Bohloolrsquos Method It is a nonparametricmethod proposed for the binarization of bad illuminateddocument images [16] In this method the morphologicalclosing operation is used to solve the background unevenillumination problem Indeed closingmay produce a reason-able estimate of the background if we use the appropriatestructuring element Experiments show that a structuringelement of size equal to twice the stroke size gives the bestresults The appropriate structuring element size is estimatedas follows A global threshold is first applied to the originalimage Then we look for the size of the largest black squarefor each pixel and we save these values in a matrix 119872 Thebiggest value of119872 in each connected set of pixels is calculatedand assigned to the other elements of the set After that the S-histogram (119891) is made starting from the matrix119872 The valueof 119891 at the point 119894 is equal to the number of elements of 119872having the value 119894 Finally we determine 119909max the greatestvalue of 119909 with 119909 satisfying

119909

sum

119894=1

119891 (119894) gt 002 times

infin

sum

119894=1

119891 (119894)

2119909

sum

119894=119909

119891 (119894) = 0 (8)

The structuring element size will be 2119909max

3 Proposed Technique

As we said earlier the global thresholding techniques aregenerally simple and fast which tend to calculate a singlethreshold in order to eliminate all background pixels andpreserve all foreground pixels Unfortunately these tech-niques are only applicablewhen the original documents are ofgood quality well contrasted and with a bimodal histogramFigure 1 shows an example

When the documents are of poor quality containingdifferent types of damages (stains transparency effects etc)with a textured background anduneven illumination orwhenthe gray levels of the foreground pixels and the gray levels ofthe background pixels are close it is not possible to find athreshold that completely separates the foreground from thebackground of the image (Figure 2)

In this case a more detailed analysis is needed and wehave recourse to local methods Local methods are moreaccurate and may be applied to variable backgrounds quitedark or with low contrast but they are very slow since thethreshold calculation based on the local neighborhood infor-mation is done for each pixel of the imageThis computationbecomes slower with larger sliding windows

To solve this problem we propose a hybrid thresholdingapproach that will be fast and at the same time effective aswell as local methods and that is achieved by combiningthe advantages of both families of binarization methods Theproposed technique uses two thresholds 119879

1and 119879

2and it

runs in two passes In the first pass a global thresholding isperformed in order to class the most of pixels of the imageAll pixels having a gray-level higher than 119879

2are removed

(becomes white) because they represent the backgroundpixels All pixels having a gray level lower than 119879

1are

considered as foreground pixels and therefore they are keptand colored in black The remaining pixels are left to thesecond pass in which they are locally binarized by combiningthe results of several local thresholding methods to select themost probable value

We detail in the following the processing steps

31 Estimation of Two Thresholds 1198791and 119879

2 The first step in

the binarization process is the calculation of the two thresh-olds119879

1and1198792 Since the thresholding purpose is to divide the


(a) Original image (b) Otsursquos thresholding result

124

Num

ber o

f pi

xels

Gray levels

(c) Corresponding histogram

Figure 2 Global thresholding of degraded image

image into two classes foreground and background and sincea single threshold is not able to accomplish this task the useof more separation thresholds seems be a perfect solution

These two thresholds are estimated from the gray levelshistogram of the original image and represent the averageintensity of the foreground and background respectively

To obtain these two thresholds we first compute a globalthreshold 119879 using a global thresholding algorithm which canbe Otsursquos algorithm Kapur or any other global algorithmIn our approach we opted for the Otsu algorithm becausethis technique has shown its efficiency and overcame otherglobal methods in several comparative studies [17 18] 119879separates the gray levels histogram of the image into twoclasses foreground and background

1198791and 119879

2are estimated from 119879 Noting 119889min the

minimum distance between the average intensity of theforeground is represented by the first half of the histogramand the average intensity of the background is represented bythe second half

1198791= 119879 minus

119889min2

1198792= 119879 +

119889min2

(9)

32 Global Image Thresholding Using 1198791and 119879

2 After the

estimation of the two thresholds 1198791and 119879

2 all pixels having

a gray level higher than 1198792are transformed into white which

eliminates most of the image background and those whosegray level is less than 119879

1are colored in black Note 119868 the

resulting image These pixels are certainly foreground pixelsThe resulting image still contains some noise but all theforeground information is preserved

33 Local Thresholding of the Remaining Pixels The pixelsunprocessed in the previous step (those with a gray levelbetween 119879

1and 119879

2) may be from the foreground and thus

must be preserved likewise they may be backgroundrsquos ornoisersquos pixels and should be removed The decision to assignthe remaining pixels to one of the two classes foreground orbackground is performed using a local process by examiningthe neighborhood of these pixels To guarantee amore correctclassification we propose to apply several local thresholdingmethods In our experiments we chose the following meth-ods Niblack Sauvola and Nick since these methods wereranked in first places in several previous comparative studies[2 19 20]

For each pixel (119909 119910) of 119868 not yet classified we calculatelocally its new binary values (0 for black and 1 for white)


obtained by applying Niblackrsquos Sauvolarsquos and Nickrsquos methodsand we obtain thus three temporary images 119868

1 1198682 and 119868

3

respectively Each one of the three local methods computesthe binary value of each remaining pixel (119909 119910) by

119868119894(119909 119910) =

0 if 119868 (119909 119910) lt LT119894(119909 119910)

255 otherwise

1 le 119894 le 3

(10)

with LT1 LT2 LT3being the local threshold computed using

Niblackrsquos Sauvolarsquos and Nickrsquos methods respectivelyThe final binary value 119868

119887(119909 119910) of each remaining pixel is

that resulted of at least two of the three methods

119868119887(119909 119910) =

0 if3

sum

119894=1

119868119894(119909 119910) lt 2

1 otherwise(11)

4 Experiments and Results

Experiments have been performed in order to estimate theperformance of our approach We applied the proposedtechnique over a large test set and compared the obtainedresults with well-known methods including global localand hybrid methods The comparison will estimate both thebinarization quality and the execution time

Firstly for parameterized methods a series of experi-ments have been performed in order to set their optimalparameter values Note PS(119898) the parameters set of a specificthresholding method 119898 For example PS(Niblack) = 119896 119908PS(Sauvola) = 119896 119908 119877 and so forth We try to find theoptimal values of PS(119898) giving the binarization results whichare closest to the ground truth images A specific range ofvalues [119886 119887] is first defined for each parameter To improvethe accuracy of the process we used a wide initial rangefor every parameter For Niblackrsquos method per example therange of the parameter 119908 is defined as [119886 119887] = [3 299] Afterthat we apply the binarization method 119898 with the differentvalues of PS(119898) from the predefined ranges on the test setdescribed in Section 41 We compare the binarization resultswith the ground truth images using the evaluation measuresdetailed in Section 42 A ranking of the obtained results isthen performed according to each measure separately Bycalculating the sum of all ranks we can infer the optimalset of parameter values as leading to the top ranks Theoptimal parameter values of the parameterized methods aresummarized in Table 1

41 Test Bases Two test sets have been used for the eval-uation of the proposed method The first is a public setcomposed of document images from the four collections pro-posed within the context of the competitions DIBCO 2009(httpusersiitdemokritosgrsimbgatDIBCO2009bench-mark) H-DIBCO 2010 (httpwwwiitdemokritosgrsimbgatH-DIBCO2010benchmark) DIBCO2011 (httputopiaduthgrsimipratikaDIBCO2011benchmark) and H-DIBCO2012 (httputopiaduthgrsimipratikaHDIBCO2012bench-

Table 1 Optimal parameters values of the parameterized methods

Binarization technique Optimal parameter valuesNiblack [11] 119896 = minus02 and 119908 = 35

Sauvola and Pietikainen [3] 119896 = 02 119877 = 128 and 119908 = 27

Nick 119896 = minus01 and 119908 = 19

Improved IGT 119899 = 50 and 119896 = 3

mark) These four collections contain a total of 50 realdocuments images (37 handwritten and 13 printed) comingfrom the collections of several libraries with the associatedground truth images All the images contain representativedegradations which appear frequently (eg variable back-ground intensity shadows smear smudge low contrastand bleed-through) Figure 3 shows some images from thesecollections

The second set of images is a synthetic collection com-posed of 150 synthetic images of documents constructedby the fusion of 15 different backgrounds and 10 binaryimages (Figure 4) The fusion is done by applying the imagemosaicing by superimposing technique for blending [21]Theidea is as follows we start with some images of documentsin black and white which represent the ground truth andwith some backgrounds extracted from old documents andwe apply a fusion procedure to get as many different imagesof old documents However Stathis et al in [22] proposedtwo different techniques for the blendingmaximum intensityand image averaging We adopt using the image averagingtechnique in order to have a more natural result

42 Evaluation Measures In addition to the execution timethe qualitative assessment is performed in terms of fivestandard evaluation measures used in DIBCO 2009 H-DIBCO 2010 DIBCO 2011 and H-DIBCO 2012 which are 119865-measure PSNR NRM MPM and DRD

Note TP TN FP and FN the true positive true negativefalse positive and false negative values respectively

421 119865-measure 119865-measure was introduced first by Chin-chor in [23]

119865-Measure = 2 times Recall times PrecisionRecall + Precision

(12)

where

Recall = TPTP + FN

Precision = TPTP + FP

(13)

422 PSNR PSNR is a similarity measure between twoimages However the higher the value of PSNR the higherthe similarity of the two images [24 25]

PSNR = 10 sdot log( 1198622

MSE) (14)


(a) (b)

(c) (d)

Figure 3 Images taken from the collections of (a) DIBCO 2009 (b) H-DIBCO 2010 (c) DIBCO 2011 and (d) H-DIBCO 2012

(a) (b) (c)

Figure 4 Image of document obtained by fusing binary image and background (a) Background from old document (b) ground truth binaryimage and (c) the resulting synthetic image

where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)

119868 and 1198682represent the two images matched 119872 and 119873

are there height and width respectively 119862 is the differencebetween foreground and background

423 NRM (Negative Metric Rate) NRM is based on thepixelwise mismatches between the ground truth and the

binarized image [26] It combines the false negative rateNRFNand the false positive rate NRFP It is denoted as follows

NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)

Contrary to 119865-measure and PSNR The better binariza-tion quality is obtained for lower NRM value


Table 2 Average results obtained on the test dataset

Execution time (ms) 119865-measure() PSNR NRM

(times10minus2)MPM(times10minus3) DRD

GlobalOtsu [6] 45875 80565 15716 883 2961 26208Kapur et al [8] 520828 81921 15692 8109 2059 23104IGT 309543 76556 14307 1011 2778 2873

LocalNiblack [11] 785041 6688 778 1147 18239 4342Sauvola and Pietikainen [3] 739142 8568 1862 795 1254 613Nick 724809 8222 19047 931 2201 527

Hybrid

Improved IGT 487943 78039 14728 1205 372 2278Gangamma and Srikanta [14] 73999157 69187 13469 2171 1543 3881Background Subtractionthresholding 1122057 68907 13432 2234 1775 4097

Tabatabaei and Bohlool [16] 75875 8582 2669 82 1183 3439Proposed Approche 141786 8744 2797 674 1025 3161

424 MPM (Misclassification PenaltyMetric) Themisclassi-fication penaltymetricMPMevaluates the binarization resultagainst the ground truth on an object-by-object basis [26]

MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894

FN and 119889119895FP denote the distance of the 119894th false negativeand the 119895th false positive pixel from the contour of the groundtruth segmentation The normalization factor 119863 is the sumover all the pixel-to-contour distances of the ground truthobject A low MPM score denotes that the algorithm is goodat identifying an objectrsquos boundary

425 DRD (Distance Reciprocal Distortion Metric) DRD isan objective distortion measure for binary document imagesand it was proposed by Lu et al in [25]Thismeasure properlycorrelates with the human visual perception and it measuresthe distortion for all the 119878 flipped pixels as follows

DRD =sum119878

119896=1DRD119896

NUBN (20)

NUBN is the number of the nonuniform 8 times 8 blocks in theGT image

DRD119896is the distortion of the 119896th flipped pixel of coordi-

nate (119909 119910) and it is calculated using a 5 times 5 normalized weightmatrix119882

119873119898 This last is defined in [25] as follows

119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2

DRD119896is given as follows

DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)

43 Results andDiscussion Theaverage results obtained overthe test images are summarized in Table 2 The final rankingof the compared methods is shown in Table 3 which alsosummarizes the partial ranks of each method according toeach evaluation measure and the sum of ranks

From Tables 2 and 3 it is clear that our proposed methodis totally ranked first and has the best performances accordingto all measures of binarization quality It exceeded localmethods such as the famous Sauvola and Pietikainenmethodwhich ranked 3rd in our experiments and even other hybridtechniques Indeed the combination of three local threshold-ing techniques enabled a more robust determination of thebinary value of each pixel by taking the most likely value

Regarding the execution time our method is very fastcompared to local methods (about 52 times more thanSauvola and Pietikainenrsquos method) which enabled us to earnabout than 98 of the execution time This is logical becauseonly a portion of the pixels (having a gray level between thetwo thresholds 119879

1and 119879

2) is locally analyzed

5 Conclusion

In this paper we tackled the problem of foregroundback-ground separation from the images of historical documents


Table 3 Final ranking of the compared methods on the test set

Rank Method Execution time(ms)

119865-measure() PSNR NRM

(times10minus2)MPM(times10minus3) DRD Sum of ranks

1 Proposed Approche 6 1 1 1 1 1 112 Tabatabaei and Bohlool [16] 10 2 2 4 2 2 223 Sauvola and Pietikainen [3] 8 3 4 2 3 4 244 Kapur et al [8] 4 5 6 3 4 6 285 Nick 7 4 3 6 5 3 286 Otsu [6] 2 6 5 5 7 7 327 IGT 1 8 8 7 6 8 388 Improved IGT 3 7 7 9 8 5 399 Gangamma and Srikanta [14] 9 9 9 10 9 9 55

10 Thresholding by BackgroundSubtraction 5 10 10 11 10 10 56

11 Niblack [11] 11 11 11 8 11 11 63

We proposed a hybrid approach of degraded documentimages binarization The proposed approach runs on twopasses Firstly a global thresholding using Otsursquos algorithm isapplied on the entire image and two different thresholds aredetermined All pixels below the first threshold are preservedand all pixels higher than the second threshold are eliminatedas they represent surely background pixels The remainingpixels are then processed locally based on their neighborhoodinformation In this step three local thresholding methodsare combined in order to obtain a more accurate decisionSince the number of pixels processed locally is very smallcompared to the total number of pixels the time required forthe binarization is reduced considerably without decreasingthe performances To validate our approach we comparedit with the state-of-the-art methods from the literature andthe obtained results on standard and synthetic collections areencouraging and confirm our approach

Conflict of Interests

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

References

[1] N Arica and F T Yarman-Vural ldquoAn overview of characterrecognition focused on off-line handwritingrdquo IEEE Transac-tions on Systems Man and Cybernetics Part C Applications andReviews vol 31 no 2 pp 216ndash233 2001

[2] K Khurshid I Siddiqi C Faure and N Vincent ldquoComparisonof Niblack inspired binarization methods for ancient docu-mentsrdquo in 16th International conference on Document Recogni-tion and Retrieval Proceedings of SPIE San Jose Calif USAJanuary 2009

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

[4] B Fernando and S Karaoglu ldquoExtreme value theory based textbinarization in documents and natural scenesrdquo in Proceedings of

the 3rd IEEE International Conference onMachineVision (ICMVrsquo10) pp 144ndash151 Hong Kong December 2010

[5] M Sezgin and B Sankur ldquoSurvey over image thresholdingtechniques and quantitative performance evaluationrdquo Journal ofElectronic Imaging vol 13 no 1 pp 146ndash168 2004

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

[7] F R D Velasco ldquoThresholding using the ISODATA clusteringalgorithmrdquo IEEE Transactions on Systems Man and Cybernet-ics vol 10 no 11 pp 771ndash774 1980

[8] J N Kapur P K Sahoo and A K C Wong ldquoA new methodfor gray-level picture thresholding using the entropy of thehistogramrdquo Computer Vision Graphics and Image Processingvol 29 no 3 pp 273ndash285 1985

[9] E Kavallieratou ldquoA binarization algorithm specialized on docu-ment images and photosrdquo in Proceedings of the 8th InternationalConference onDocument Analysis and Recognition pp 463ndash467September 2005

[10] J Bernsen ldquoDynamic thresholding of grey-level imagesrdquo inProceedings of the 8th International Conference on PatternRecognition pp 1251ndash1255 Paris France 1986

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

[12] T Sari A Kefali and H Bahi ldquoAn MLP for binarizing imagesof old manuscriptsrdquo in Proceedings of the 13th InternationalConference on Frontiers in Handwriting Recognition (ICFHRrsquo12) pp 247ndash251 Bari Italy September 2012

[13] E Kavallieratou and S Stathis ldquoAdaptive binarization of histor-ical document imagesrdquo in Proceedings of the 18th InternationalConference on Pattern Recognition (ICPR rsquo06) vol 3 pp 742ndash745 Hong Kong August 2006

[14] B Gangamma and M K Srikanta ldquoEnhancement of degradedhistorical kannada documentsrdquo International Journal of Com-puter Applications vol 29 no 11 pp 1ndash6 2011

[15] G Leedham C Yan K Takru J H N Tan and LMian ldquoCom-parison of Some thresholding algorithms for textbackgroundsgmentation in difficult document imagesrdquo in Proceedings ofthe 7th International Conference on Document Analysis andRecognition pp 859ndash864 2003


[16] S A Tabatabaei and M Bohlool ldquoA novel method for bina-rization of badly illuminated document imagesrdquo in Proceedingsof the 17th IEEE International Conference on Image Processing(ICIP rsquo10) pp 3573ndash3576 Hong Kong China September 2010

[17] P K Sahoo S Soltani and A K C Wong ldquoA survey ofthresholding techniquesrdquoComputer Vision Graphics and ImageProcessing vol 41 no 2 pp 233ndash260 1988

[18] M A Ramırez-Ortegon and R Rojas ldquoUnsupervised eval-uation methods based on local gray-intensity variances forbinarization of historical documentsrdquo in Proceedings of the 20thInternational Conference on Pattern Recognition (ICPR rsquo10) pp2029ndash2032 IEEE Computer Society Istanbul Turkey August2010

[19] A J O Trier ldquoGoal-directed evaluation of binarization meth-odsrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 17 no 12 pp 1191ndash1201 1995

[20] A Kefali T Sari and M Sellami ldquoEvaluation of severalbinarization techniques for old Arabic documents imagesrdquoin Proceedings of International Symposium on Modelling andImplementation of Complex Systems Constantine Algeria 2010

[21] L G Brown ldquoA survey of image registration techniquesrdquo ACMComputing Surveys vol 24 no 4 pp 325ndash376 1992

[22] P Stathis E Kavallieratou and N Papamarkos ldquoAn evaluationsurvey of binarization algorithms on historical documentsrdquoin Proceedings of the 19th International Conference on PatternRecognition pp 742ndash745 December 2008

[23] N Chinchor ldquoMUC-4 evaluation metricsrdquo in Proceedings of the4th Message Understanding Conference pp 22ndash29 1992

[24] B Gatos K Ntirogiannis and I Pratikakis ldquoDIBCO 2009document image binarization contestrdquo International Journal onDocument Analysis and Recognition vol 14 no 1 pp 35ndash442011

[25] H Lu A C Kot and Y Q Shi ldquoDistance-reciprocal distortionmeasure for binary document imagesrdquo IEEE Signal ProcessingLetters vol 11 no 2 pp 228ndash231 2004

[26] J Aguilera H Wildenauer M Kampel M Borg D Thirdeand J Ferryman ldquoEvaluation of motion segmentation qualityfor aircraft activity surveillancerdquo in Proceedings of the 2ndJoint IEEE International Workshop on Visual Surveillance andPerformance Evaluation of Tracking and Surveillance (VS-PETSrsquo05) pp 293ndash300 October 2005

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



methods in contrast combine global and local informationfor segmenting the image

In this paper we propose a new hybrid thresholdingtechnique for binarizing images of historical documentsThe proposed approach uses a mixture of thresholdingmethods and it combines the advantages of the two familiesof techniques the computation speed and the efficiencyThe remainder of this paper is organized as follows InSection 2 we present some existing binarization methodsThen in Section 3 we describe the proposed approach Theexperiments performed and the results will be shown inSection 4 before concluding

2 State of the Art

According to [4] existing methodologies for image bina-rization may be divided under two main strategies groupingbased and thresholding based Thresholding based methodsuse global or local threshold(s) to separate the text from thebackground In grouping based methods we distinguish twocategories region based grouping and clustering based group-ing methods Region based grouping methods are mainlybased on spatial-domain region growing or on splitting andmerging Also clustering based grouping methods are basedon classification of intensity or color values as a functionof a homogeneity criterion However several techniqueshave been employed to reach this classification K-meansalgorithm artificial neural networks and so forth

Sezgin and Sankur [5] established a classification ofbinarization methods according to the information that theyexploit in 6 categories

(i) Histogram-based methods the methods of this classperform a thresholding based on the form of thehistogram

(ii) Clustering-based methods these methods assign theimage pixels to one of the two clusters object andbackground

(iii) Entropy-based methods these algorithms use theinformation theory to obtain the threshold

(iv) Object attribute-based methods they find a thresh-old value based on some similarity measurementsbetween original and binary images

(v) Spatial binarization methods they find the optimalthreshold value taking into account spatial measures

(vi) Locally adaptive methods these methods aredesigned to give a new threshold for every pixelSeveral kinds of adaptive methods exist We findmethods based on local gray range local variationand so forth

Wepresent in this section some binarizationmethods themost frequently cited in the literature and we consider onlythresholding based methods

21 Global Methods Note 119868 the grayscale image of whichthe intensities vary from 0 (black) to 1 (white) and 119867 its

histogram of intensities The number of pixels having a graylevel 119894 is denoted as119867(119894)

211 Otsursquos Method Otsursquos method [6] tries to find thethreshold 119879 which separates the gray-level histogram inan optimal way into two segments (which maximize theintersegments variance or whichminimize the intrasegmentsvariance) The calculation of the interclasses or intraclassesvariances is based on the normalized histogram 119867

119899=

[119867119899(0) sdot sdot sdot 119867

119899(255)] of the image where sum119867

119899(119894) = 1

The interclasses variance for each gray level 119905 is given by

119881inter = 1199021(119905) times 119902

2(119905) times [120583

1(119905) minus 120583

2(119905)]2 (1)

such that

1205831(119905) =

1

1199021(119905)

119905minus1

sum

119894=0

119867119899(119894) times 119894

1205832(119905) =

1

1199022(119905)

255

sum

119894=119905

119867119899(119894) times 119894

1199021(119905) =

119905minus1

sum

119894=0

119867119899(119894) 119902

2(119905) =

255

sum

119894=119905

119867119899(119894)

(2)

212 ISODATA Method Thresholding using ISODATA [7]consists in finding a threshold by separating iteratively thegray-level histogram into two classes with the a prioriknowledge of the values associated with each class Thismethod starts by dividing the interval of nonnull values ofthe histogram into two equidistant parts and next we take119898

1

and 1198982as the arithmetic average of each class Repeat until

convergence the calculation of the optimal threshold 119879 as theclosest integer to (119898

1+1198982)2 and update the two averages119898

1

and1198982

213 Kapur et alrsquos Method Kapur et alrsquos method [8] isan entropy based method which takes into account theforeground likelihood distribution 119875

119891and the background

likelihood distribution (119875119887= 1 minus 119875

119891) in the determination of

the division entropy The binarization threshold 119879 is chosenfor which the value 119864 = 119864

119891+ 119864119887is maximal such that

119864119891= minus

119905

sum

119894=0

119875119894

119875119891

times log(119875119894

119875119891

)

119864119887= minus

255

sum

119894=119905+1

119875119894

1 minus 119875119891

times log(119875119894

1 minus 119875119891

)

(3)

where 119875119894is the occurrence probability of the gray level 119894 in the

image and 119875119891= sum119905

119894=0119875119894

214 Iterative Global Thresholding (IGT) The proposedmethod selects a global threshold to the entire image basedon an iterative procedure [9] At each iteration 119894 the followingsteps are performed

(a) calculating the average gray level (119879119894) of the image

(b) subtracting 119879119894from all pixels of the image






119879 (119909 119910) =

119885low + 119885high

2

(4)




119879 (119909 119910) = 119898 + 119896 times 120590 (5)




119877)) (6)



119879 (119909 119910) = 119898 + 119896radic(sum1199012

119894minus 1198982

)

NP(7)





23 Hybrid Methods






Num

ber o

f pi

xels

Gray levels

133

(b) Its histogram




119909

sum

119894=1

119891 (119894) gt 002 times

infin

sum

119894=1

119891 (119894)

2119909

sum

119894=119909

119891 (119894) = 0 (8)







1and 119879

2and it


2are removed


1are




2 The first step in





124

Num

ber o

f pi

xels

Gray levels






1198791and 119879



1198791= 119879 minus

119889min2

1198792= 119879 +

119889min2

(9)


2 After the


2 all pixels having






1and 119879






1 1198682 and 119868

3


119868119894(119909 119910) =

0 if 119868 (119909 119910) lt LT119894(119909 119910)

255 otherwise

1 le 119894 le 3

(10)





119868119887(119909 119910) =

0 if3

sum

119894=1

119868119894(119909 119910) lt 2

1 otherwise(11)








Nick 119896 = minus01 and 119908 = 19

Improved IGT 119899 = 50 and 119896 = 3







(12)

where

Recall = TPTP + FN


(13)



MSE) (14)


(a) (b)

(c) (d)


(a) (b) (c)


where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)





NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)








Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














119879 (119909 119910) =

119885low + 119885high

2

(4)




119879 (119909 119910) = 119898 + 119896 times 120590 (5)




119877)) (6)



119879 (119909 119910) = 119898 + 119896radic(sum1199012

119894minus 1198982

)

NP(7)





23 Hybrid Methods






Num

ber o

f pi

xels

Gray levels

133

(b) Its histogram




119909

sum

119894=1

119891 (119894) gt 002 times

infin

sum

119894=1

119891 (119894)

2119909

sum

119894=119909

119891 (119894) = 0 (8)







1and 119879

2and it


2are removed


1are




2 The first step in





124

Num

ber o

f pi

xels

Gray levels






1198791and 119879



1198791= 119879 minus

119889min2

1198792= 119879 +

119889min2

(9)


2 After the


2 all pixels having






1and 119879






1 1198682 and 119868

3


119868119894(119909 119910) =

0 if 119868 (119909 119910) lt LT119894(119909 119910)

255 otherwise

1 le 119894 le 3

(10)





119868119887(119909 119910) =

0 if3

sum

119894=1

119868119894(119909 119910) lt 2

1 otherwise(11)








Nick 119896 = minus01 and 119908 = 19

Improved IGT 119899 = 50 and 119896 = 3







(12)

where

Recall = TPTP + FN


(13)



MSE) (14)


(a) (b)

(c) (d)


(a) (b) (c)


where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)





NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)








Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Num

ber o

f pi

xels

Gray levels

133

(b) Its histogram




119909

sum

119894=1

119891 (119894) gt 002 times

infin

sum

119894=1

119891 (119894)

2119909

sum

119894=119909

119891 (119894) = 0 (8)







1and 119879

2and it


2are removed


1are




2 The first step in





124

Num

ber o

f pi

xels

Gray levels






1198791and 119879



1198791= 119879 minus

119889min2

1198792= 119879 +

119889min2

(9)


2 After the


2 all pixels having






1and 119879






1 1198682 and 119868

3


119868119894(119909 119910) =

0 if 119868 (119909 119910) lt LT119894(119909 119910)

255 otherwise

1 le 119894 le 3

(10)





119868119887(119909 119910) =

0 if3

sum

119894=1

119868119894(119909 119910) lt 2

1 otherwise(11)








Nick 119896 = minus01 and 119908 = 19

Improved IGT 119899 = 50 and 119896 = 3







(12)

where

Recall = TPTP + FN


(13)



MSE) (14)


(a) (b)

(c) (d)


(a) (b) (c)


where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)





NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)








Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











124

Num

ber o

f pi

xels

Gray levels






1198791and 119879



1198791= 119879 minus

119889min2

1198792= 119879 +

119889min2

(9)


2 After the


2 all pixels having






1and 119879






1 1198682 and 119868

3


119868119894(119909 119910) =

0 if 119868 (119909 119910) lt LT119894(119909 119910)

255 otherwise

1 le 119894 le 3

(10)





119868119887(119909 119910) =

0 if3

sum

119894=1

119868119894(119909 119910) lt 2

1 otherwise(11)








Nick 119896 = minus01 and 119908 = 19

Improved IGT 119899 = 50 and 119896 = 3







(12)

where

Recall = TPTP + FN


(13)



MSE) (14)


(a) (b)

(c) (d)


(a) (b) (c)


where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)





NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)








Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











1 1198682 and 119868

3


119868119894(119909 119910) =

0 if 119868 (119909 119910) lt LT119894(119909 119910)

255 otherwise

1 le 119894 le 3

(10)





119868119887(119909 119910) =

0 if3

sum

119894=1

119868119894(119909 119910) lt 2

1 otherwise(11)








Nick 119896 = minus01 and 119908 = 19

Improved IGT 119899 = 50 and 119896 = 3







(12)

where

Recall = TPTP + FN


(13)



MSE) (14)


(a) (b)

(c) (d)


(a) (b) (c)


where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)





NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)








Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)

(c) (d)


(a) (b) (c)


where

MSE =

sum119872

119909=1sum119873

119910=1(119868 (119909 119910) minus 119868

2(119909 119910))

2

119872119873

(15)





NRM =NRFN +NRFP

2(16)

with

NRFN =FN

FN + TP NRFP =

FPFP + TN

(17)








Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















Hybrid




MPM =MPFN +MPFP

2 (18)

where

MPFN =sum

FN119894=1

119889119894

FN119863

MPFP =sum

FP119895=1

119889119895

FP

119863

(19)

119889119894



DRD =sum119878

119896=1DRD119896

NUBN (20)





119882119873119898

(119894 119895) =119882119898(119894 119895)

sum119898

119894=1sum119898

119895=1119882119898(119894 119895)

(21)

such that

119882119898(119894 119895) =

0 for 119894 = 119894119862 119895 = 119895

119862

1

radic(119894 minus 119894119862)2

+ (119895 minus 119895119862)2

otherwise

(22)

With119898 = 5 and 119894119862= 119895119862= (1 + 119898)2


DRD119896=

2

sum

119894=minus2

2

sum

119895=minus2

1003816100381610038161003816119868GT (119909 + 119894 119910 + 119895) minus 119868119861 (119909 119910)1003816100381610038161003816

times 119882119873119898

(119894 + 2 119895 + 2)

(23)




1and 119879


5 Conclusion









11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















11 Niblack [11] 11 11 11 8 11 11 63




References































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation























RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









research article text extraction from historical document

Documents