research article an improved saliency detection approach ...apsaras with di erent historical periods...

Research ArticleAn Improved Saliency Detection Approach for Flying Apsaras inthe Dunhuang Grotto Murals China

Zhong Chen1 Shengwu Xiong1 Qingzhou Mao23 Zhixiang Fang23 and Xiaohan Yu1

1Wuhan University of Technology School of Computer Science and Technology 122 Luoshi Road Wuhan 430070 China2Wuhan University State Key Laboratory for Information Engineering in Surveying Mapping and Remote Sensing129 Luoyu Road Wuhan 430079 China3Wuhan University Engineering Research Center for Spatio-Temporal Data Smart Acquisition and ApplicationMinistry of Education of China 129 Luoyu Road Wuhan 430079 China

Correspondence should be addressed to Qingzhou Mao qzhmaowhueducn

Received 20 September 2014 Revised 5 January 2015 Accepted 23 January 2015

Academic Editor Chong Wah Ngo

Copyright copy 2015 Zhong Chen 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

Saliency can be described as the ability of an item to be detected from its background in any particular scene and it aims toestimate the probable location of the salient objects Due to the salient map that computed by local contrast features can extractand highlight the edge parts including painting lines of Flying Apsaras in this paper we proposed an improved approach based ona frequency-tuned method for visual saliency detection of Flying Apsaras in the Dunhuang Grotto Murals China This improvedsaliency detection approach comprises three important steps (1) image color and gray channel decomposition (2) gray featurevalue computation and color channel convolution (3) visual saliency definition based on normalization of previous visual saliencyand spatial attention function Unlike existing approaches that rely on many complex image features this proposed approach onlyused local contrast and spatial attention information to simulate humanrsquos visual attention stimuliThis improved approach resultedin a much more efficient salient map in the aspect of computing performance Furthermore experimental results on the dataset ofFlying Apsaras in the Dunhuang Grotto Murals showed that the proposed visual saliency detection approach is very effective whencompared with five other state-of-the-art approaches

1 Introduction

In Chinese traditional culture Buddhism (particularly in thewest of China) has played a prominent role in Buddhismhistory This link is conveyed via the Dunhuang GrottoMurals which have been of significant value to culturalstudies across various Chinese dynasties Protecting andmaintaining the cultural integrity of the Dunhuang GrottoMurals have been a tremendous challenge for researchersDigitalization comprehensive preservation and virtual inter-action have been adopted as the key approaches to protectingand propagating the culture embodied in the DunhuangGrotto Murals

Flying Apsaras are a significant symbol of the Dunhuangart and contain a lot of cultural information and connota-tions which have been composed by a wide variety of peoplenamely streamers clouds totem and so on In addition

because of the age diverse colors and complex compositionof the images in Flying Apsaras extracting the outlines accu-rately is of great significance in analyzing the painting styles ofthe Dunhuang Grotto Murals in different Chinese dynastiesFurthermore due to the multioverlap phenomenon existingin some images of the Dunhuang Grotto Murals that is laterworks were painted directly on the early works analyzingthe cultural meanings of these images is even more difficultFortunately one approach called saliency detection canbe used to analyze the different painting styles of FlyingApsaras with different historical periods in the DunhuangGrotto Murals The relationship between Flying Apsaras andvisual saliency can be closely connected by two importantaspects on the one hand visual saliency can represent theinitial state of these images without coloring and saliencydetection of Flying Apsaras in the Dunhuang Grotto Muralsis of great historical value for historians in helping determine

Hindawi Publishing CorporationAdvances in MultimediaVolume 2015 Article ID 625915 11 pageshttpdxdoiorg1011552015625915

2 Advances in Multimedia

the original state of the images in the Dunhuang GrottoMurals on the other hand the final salient map of FlyingApsaras that computed by local contrast features can extractand highlight the edge parts with striking local contrast theseparts including the painting lines have played a critical rolein analyzing significant painting styles of Flying Apsaras Inaddition visual saliency detection can be used to extractthe salient regions of Flying Apsaras and these regions canrepresent the remarkable paintingsrsquo main thought designand plot

One of the most challenging problems with perceptionis information overload This overload presents a challengewhen trying to obtain useful information from images dueto the semantic gap between low level features and high levelsemantic concepts [1] Visual attention is one of the primaryfeatures of human visual system (HVS) to derive importantand compact information from the natural scenes [2] Andits mechanism enables a reduction of the redundant data thatbenefits perception during the selective attention process Inthe image processing field saliency detection is the processof detecting the interesting visual information Furthermorethe selected stimuli need to be prioritized with the mostrelevant being processed first and the less important oneslater and it leads to a sequential treatment of different partsof the visual scenes Thus saliency detection has been widelyused in discovering regions of interests predicting humanvisual attention image quality assessment and many otherapplications [3] The saliency detection approach is used inmany fields including object detection object recognitionimage enhancement image rendering image compressionand video summarization [4]

Many different approaches to visual saliency detectionhave been proposed by a variety of researchers from dif-ferent research backgrounds As pioneers Koch Ullmanand Poggio [5 6] proposed a very influential Bioinspiredmodel and Winner-Take-All selected mechanism to simulatevisual system of humans Itti et al [7] derived bottom-up andtop-down visual saliency using center-surround differencesvia multiscale features integration Harel et al [8] proposeda new bottom-up visual saliency detection model graph-based visual saliency (GBVS) to form activation maps oncertain feature channels and normalize them with highlight-ing conspicuity and combining with other maps Liu et al[9] used the conditional random field (CRF) approach toeffectively combine a set of novel features including multi-scale contrast center-surround histogram and color spatialdistribution to describe salient object locally regionallyand globally Goferman et al [10] proposed a new type ofsaliency model context-aware saliency which is based onfour principles observed in the psychological literature locallow-level features global considerations visual organizationrules and high-level factors Zhai and Shah [11] developeda fast approach for computing pixel-level salient maps usingcolor histograms of images and a dynamic fusion techniqueapplied to combine the temporal and spatial salient mapsHou and Zhang [12] proposed a spectral residual approachfor visual saliency detection to construct the correspondingsalient maps in the spatial domain and extract the spectralresidual of an image by analyzing the log-spectrum of it

This is replaced by the phase spectrum of the Fouriertransform in Gopalakrishnan et al [13] because it is moreeffective and computationally efficient Achanta et al [14]combined low-level features of color and luminance to detectsalient regions of images which was called a frequency-tunedmethod More recently Yan et al [15] employed a multilayerapproach to analyze saliency cues and the final salient mapwas produced in a hierarchical model Siva et al [16] usedan unsupervised learning approach that annotated objects ofinterest in images to capture the salient maps

Due to employing only a few low-level features of imagesthe visual saliency detection results of the frequency-tunedmethod proposed by Achanta et al [14] are not very sat-isfactory In this paper an alternative approach histogramcombined with image average and Gaussian blur (HIG) isproposed as an improvement to the frequency-tunedmethodHIG combines features of color and luminance with thegray histogram information and uses the Euclidean distancesbetween local features and global average features to definevisual saliency as local contrast at pixel level And thefinal visual saliency definition is based on normalization ofprior visual saliency and spatial attention function Unlikeexisting approaches that focus on multiple image featuresthe proposed approach only focuses on contrast and spatialattention information from the simulation of HVS and theselocal contrast features have played a key role in highlightingthe edge parts including painting lines of Flying ApsarasThisallows for more efficient computation of salient maps andvery effective performancewhen compared to five other state-of-the-art approaches These five state-of-the-art approachesof saliency detection are Zhai and Shah [11] Hou and Zhang[12] Achanta et al [14] and Cheng et al [17] hereby referredto as LC SR IG HC and RC respectively The choice ofthese algorithms was based on their frequency of citation inliterature popularity and variety

The rest of this paper is organized as follows Section 2provides an overview of the proposed saliency detectionapproach followed by a more detailed description Section 3presents the experimental validations and comparison ofresults finally Section 4 contains the conclusion and recom-mendations for future works

2 The Proposed Saliency Detection Approach

An overview of the improved saliency detection approachproposed for Flying Apsaras in the Dunhuang Grotto Muralsin this paper is shown in Figure 1 This improved approachcomprises the following three important steps (1) imagecolor and gray channel decomposition (2) gray feature valueand color channel convolution (3) visual saliency definitionbased on normalization of prior visual saliency and spatialattention function Using these three steps the final salientmap is determined from the input image that is shown as anexample in Figure 1

21 Decomposition of Color and Gray Components fromthe Input Images It has been widely recognized amongresearchers that HVS does not equally process all

Advances in Multimedia 3

Lab image (L)

Lab image (A)

Lab image (B)

Salient map

Gauss convolution

Gre

y va

lue

Grey features value

spac

e

Col

or sp

ace

conv

ersio

n

Grey image Input image

Final saliency definitionVisual saliency definition

Grey average value Lab average value

Normaliz-

Spatialattention

H(i j) I(i j)Local features

I120583

Global average features

S(i j) = SN(i j) middot F(i j)SP(i j) = I(i j) minus I120583

|H(i j) minus

ation

I120583 = lfloorlceil

rfloorrceil

I(i j) =

L(i j)

A(i j)

B(i j)lfloorlceilrfloorrceil

+

=

m

sumi=1

n

sumj=1

H(i j)(m times n)

H(i j) = 1

21205871205902eminus((iminus(m2))

2+(jminusn2))

22120590

2)

Avg LAvg AAvg B

Avg H|

Avg H

Avg H

Figure 1 An overview of our saliency detection approach for Flying Apsaras in the Dunhuang Grotto Murals

the information available to an observer Color contrasthas a significant impact on the ability to detect salientregions in images and has been highlighted in many previousworks For example Osberger and Rohaly [18] suggest that astrong influence occurs when the color of a region is distinct

from its background and some particular colors (eg redyellow and blue) attract our attention more than others

For an RGB image the three color channels are correlatedlinearly In addition because the value of each color channelis not related to the representation of stimulus intensity


it is difficult to compute the difference in color intensityUsing a variety of color transformation formulas other colorspaces (eg HSI Lab and LUV which are specified by theInternational Commission on Illumination (CIE)) can bederived from the RGB color space The CIErsquos Lab colorspace is closest to human visual perception and perceptualuniformity where 119871 represents luminance information andmatches human perception of lightness closely 119886 and 119887represent chromatic values and the transformation formulacan be shown in the following [19]

119883 = 0490 times 119877 + 0310 times 119866 + 0200 times 119861


119885 = 0010 times 119866 + 0990 times 119861

119871 = 116119891 (119884) minus 16

119886 = 500 sdot [119891 (119883

0982) minus 119891 (119884)]

119887 = 200 sdot [119891 (119884) minus 119891(119885

1183)]

(1)

where

119891 (119883) =

7787119883 + 0138 119883 le 0008856

11988313 119883 gt 0008856

(2)

In this paper we utilize this nonlinear relational mapping ofcolor channels 119871 119886 and 119887 and gray histogram informationto imitate the nonlinear response of the human eyes Inaddition CIErsquos Lab is the most widely used and efficientimage features for visual saliency detection algorithms thatis researches in Itti et al [7] Achanta et al [20] Achanta etal [14] Goferman et al [10] Cheng et al [17] and Yang et al[21]

22 The Frequency-Tuned Saliency Detection Approach Vis-ual saliency refers to the ability of a vision system (humanor machine) to select from an image a certain subset of visualinformation for further processingThismechanism serves asa filter to select only the relevant information in the context ofthe current behaviors or tasks to be processed while ignoringother irrelevant information [4] Achanta et al [14] define thefive key functions for a saliency detector as being able to (1)emphasize the largest salient objects (2) uniformly highlightwhole salient regions (3) establish well-defined boundariesof salient objects (4) disregard high frequencies arisingfrom texture noise and blocking artifacts and (5) efficientlyoutput full resolution salient maps These functions can beused when comparing alternative visual saliency detectionapproaches

From the perspective of image frequency IG approach[14] suggests that the input image can be divided into theconstituent low frequency and high frequency parts of thefrequency domainTheoverall parts of structure informationsuch as contours and basic compositions of the originalimage are reflected in the low frequency parts while details

such as texture andnoise are reflected in high frequency partsIn saliency detection parts the overall structure informationcontained in low frequency parts has been widely used tocompute visual saliency

Let120596119897119888and120596

ℎ119888represent theminimum frequency and the

maximum frequency in the frequency domain respectivelyOn the one hand in order to accurately highlight theoverall information representing potential salient objects120596119897119888should be low and it is helpful to highlight the entire

object consistently and uniformly On the other hand inorder to maintain the rich semantic information containedin any potential salient objects 120596

ℎ119888should be high while

remaining cognizant of the fact that the highest elements inthe frequency domain should be discarded for they canmostlikely be attributed to texture or noise parts

Achanta et al [14] chose the Difference of Gaussian(DoG) filter (3) for band-pass filtering to capture frequencylimits 120596

119897119888and 120596

ℎ119888 This DoG filter has also been used for

interest point detection [22] and saliency detection [7 8]TheDoG filter is defined as

DoG (119909 119910) = 119866 (119909 119910 1205901) minus 119866 (119909 119910 120590

2)

=1

2120587[1

1205902

1

119890minus(1199092+1199102)21205902

1 minus1

1205902

2

119890minus(1199092+1199102)21205902

2]

(3)

where 1205901and 120590

2(1205901lt 1205902) are the standard deviations of the

Gaussian functionA DoG filter is a simple band-pass filter and its pass-

band width is controlled by the ratio 12059011205902 The DoG filter

is widely used in edge detection since it closely and efficientlyapproximates the Laplacian of Gaussian (LoG) filter cited asthe most satisfactory function for detecting intensity changeswhen the standard deviations of the Gaussian function are inthe ratio 120590

11205902= 1 16 [23] If we define 120590

1= 120588120590 and 120590

2= 120590

namely 120588 = 12059011205902 then the bandwidth of the DoG can be

determined by 120588 Because a single DoG operation cannotcover the bandwidth [120596

119897119888 120596ℎ119888] several narrowband-passDoG

filters are combined and it is found that a summation overDoG with standard deviations can cover the bandwidth bythe ratio 120588 the formula is given by

119873minus1

sum

119899=0

[119866 (119909 119910 120588119899+1120590) minus 119866 (119909 119910 120588

119899120590)]

= 119866 (119909 119910 120588119873120590) minus 119866 (119909 119910 120590)

(4)

for an integer119873 ge 0 which is the difference of two Gaussianswhose standard deviations can have any ratio119870 = 120588119873That iswe can obtain the combined result of applying several band-pass filters by choosing a DoG with a relatively large 119870 Inorder to make the bandwidth [120596

119897119888 120596ℎ119888] as large as possible119870

must be large In the extreme case119873 = infin then 119866(119909 119910 120588119873120590)is actually the average vector of the entire imageThis processillustrates how the salient regions will be fully covered andnot just focused on edges or in the center of the regions [14]


Themain steps of IG approach are as follows (1) the inputRGB image 119868 is transformed to a CIErsquos Lab color space (2) thesalient map 119878 for the original image 119868 is computed by

119878 (119909 119910) =10038171003817100381710038171003817119868120596ℎ119888

(119909 119910) minus 119868120583

10038171003817100381710038171003817 (5)

where 119868120583

is the arithmetic mean image feature vector119868120596ℎ119888

(119909 119910) is a Gaussian blurred version of the original imageusing a 5 times 5 separable binomial kernel sdot is the 119871

2norm

and 119909 and 119910 are the pixel coordinatesIn conclusion IG approach is a very simple visual saliency

method to implement and only requires use of a Gaussianblur and image averaging process From the point of view ofimplementation IG approach is based on a global contrastmethod and the global information is the mean featurevector of the original image

23 The Improved Saliency Detection Approach The saliencyof an input image can be described as the probability thatan item distinct from its background will be detectedand output to a salient map in the form of an intensitymap The saliency detection algorithm aims to estimate theprobable location of the salient object A larger intensityvalue at a pixel indicates a higher saliency value whichmeans the pixel most probably belongs to the salient objectIn Achanta et al [14] features of color and luminancewere exploited to define visual saliency Specifically the 119871

2

norm between the corresponding image pixel vector valuein the Gaussian blurred version and the mean image featurevector of the original image In many natural scenes par-ticularly in the three widely used datasets including Brucersquos[24] eye tracking dataset (BET) Juddrsquos [25] eye trackingdataset (JET) and Achantarsquos [14] human-annotated dataset(AHA) the frequency-tuned saliency detection algorithmperforms excellently But like many other saliency detectionapproaches it is not effective for images of Flying Apsaras inthe DunhuangGrottoMuralsThis is probably due to the richcontent of the Dunhuang Grotto Murals which presents abig challenge to distinguishing foregrounds from the muralrsquosbackgrounds

In order to maintain the complete structural informationof salient objects as well as considering features of colorand luminance this paper also takes the gray histograminformation of the original image and human visual attentionstimuli into account and proposes an improved saliencydetection approach for Flying Apsaras in the DunhuangGrotto Murals The gray feature contains rich luminanceinformation and can be used to detect the salient objects of animage In this paper after converting the original image intogray-scale image we can define the gray feature value119867(119894 119895)of pixel 119901(119894 119895) as

119867(119894 119895) =1

21205871205902119890minus((119894minus1198982)

2+(119895minus1198992)

2)21205902

(6)

where 120590 is the normalized variance of Gaussian function and119898 and 119899 represent the height and width of the input image 119868

respectively Thus the mean value Avg119867 of gray feature canbe expressed by

Avg119867 =sum119898

119894=1sum119899

119895=1119867(119894 119895)

119898 times 119899 (7)

where 119898 times 119899 represents the total number of pixels of theoriginal image

Meanwhile after converting the input image inRGB colorspace into Lab color space we can obtain the luminance value119871(119894 119895) and color feature values 119860(119894 119895) and 119861(119894 119895) of a pixel atthe position (119894 119895) Then using (3) to calculate the Gaussianburred version 119868

120596ℎ119888

(119894 119895) = (119866119884119871(119894 119895) 119866119884119860(119894 119895) 119866119884119861(119894 119895))119879

of the original image as expressed in Achantarsquos researchthe salient regions will be fully covered and not just behighlighted on their edges or in the center of the salientregions Then the arithmetic mean image feature vector 119868

120583=

(Avg 119871Avg119860Avg119861)119879 of the input image can be determinedby

Avg 119871 =sum119898

119894=1sum119899

119895=1119871 (119894 119895)

119898 times 119899 Avg119860 =

sum119898

119894=1sum119899

119895=1119860 (119894 119895)

119898 times 119899

Avg119861 =sum119898

119894=1sum119899

119895=1119861 (119894 119895)

119898 times 119899

(8)

Combining features of color and luminance with the grayhistogram information this paper exploits the visual contrastprinciple to define visual saliency as local contrast based onpixel level We consider that pixels with significant differencecompared with the mean vector are part of the salient regionsand assume that pixels with values close to the mean vectorare part of the background Therefore the saliency 119878

119875(119894 119895)

of a pixel at location (119894 119895) is characterized by the differencebetween it and the mean vector namely

119878119875(119894 119895) =

10038171003817100381710038171003817119868120596ℎ119888

(119894 119895) minus 119868120583

10038171003817100381710038171003817+1003816100381610038161003816119867 (119894 119895) minus Avg119867

1003816100381610038161003816 (9)

where sdot is the 1198712norm and |sdot| is the absolute value function

Because the final salient map is actually a gray imagein order to optimize display of the final salient map it canbe processed by a gray-scale transformation based on somenormalized functions In this paper the normalized salientmap 119878

119873(119894 119895) can be further computed by

119878119873(119894 119895) = 255 times

119878119875(119894 119895) minus 119878min119878max minus 119878min

(10)

where 119878max and 119878min denote the maximum and minimumvalues of the salient map 119878

119875(119894 119895) respectively Normalization

ensures these values of 119878119873(119894 119895) have a unified range from 0

to 255 for different image locations which is convenient forvisualization of final output

Generally when an observer views an image he or she hasa tendency to stare at the central location without makingeye movements This phenomenon is referred to as centraleffect [26] In Zhang et al [26] when searching the scenes fora conspicuous target fixation distributions are shifted from


the image center to the distributions of image features Basedon the center shift mechanism in this paper we define aspatial attention function 119865(119894 119895) a weight decay functionto reflect the spatial attention tendency of human eyes asfollows

119865 (119894 119895) =119871 minus 2119863 (119894 119895)

119871 + 2119863 (119894 119895) (11)

where 119863(119894 119895) denotes the distance between pixel at location(119894 119895) and the center of a given image and 119871 is the length ofthe diagonal of the input image For any pixel at location(119894 119895) 119863(119894 119895) isin [0 1198712] thus 119865(119894 119895) isin [0 1] and 119865(119894 119895) is amonotonically decreasing function with the variable119863(119894 119895)

In summary the definition of the spatial attention func-tion is reasonable since it does not exclude the possibilityfor the pixels located at the boundary of an image to benoticed [26] In order to integrate the extracted features andinteraction effects of human observations that are measuredby the spatial attention function 119865(119894 119895) the final salient mapin the proposed approach is evaluated by

119878 (119894 119895) = 119865 (119894 119895) sdot 119878119873(119894 119895) (12)

This improved saliency detection approach of combininggray histogram information features of color and luminanceand the humanrsquos spatial visual attention function for solvingvisual saliency detection problem of the Flying Apsaras in theDunhuang Grotto Murals can be expressed by Algorithm 1

Algorithm 1 The improved saliency detection approach forFlying Apsaras in the Dunhuang Grotto Murals China is asfollows

Input Images of Flying Apsaras in the DunhuangGrotto Murals ChinaOutput Visual saliency detection maps of images inthe Dunhuang Grotto Murals ChinaProcess(1) Convert the original image into a grayscale imageand then compute the gray feature value 119867(119894 119895) ofpixel 119901(119894 119895) according to the gray feature functionaccording to (6) Obtain themean valueAvg119867 of grayfeature simultaneously(2) Convert the original RGB color space image intoCEIrsquos Lab color space format to obtain three differentproperty values 119871(119894 119895) 119860(119894 119895) and 119861(119894 119895) of threedifferent color channels 119871 119886 and 119887(3) Use a 5times5 separable binomial kernel of pixel at theposition to obtain the Gaussian blurred version 119868

120596ℎ119888

of the original image and obtain the arithmetic meanimage feature vector 119868

120583of input image simultaneously

(4) Define the new saliency detection formula shownin (9) and use (10) to export the normalized salientmap 119878

119873(119894 119895)

(5) Integrate the spatial attention function 119865(119894 119895) tothe normalized salient map 119878

119873(119894 119895) to obtain the final

definition of visual saliency detection 119878(119894 119895)(6) Export the final visual saliency detection maps ofthe original images

3 Experimental Results

In order to evaluate the performance of this proposed visualsaliency detection approach an image dataset of FlyingApsaras in the Dunhuang Grotto Murals was establishedfrom Zheng and Tairsquos [27] book To prove the effectivenessof the proposed approach this paper compares experimentalresults obtained using the proposed approach with thoseobtained when applying five other state-of-the-art algo-rithms

31 The Dataset of Flying Apsaras in the Dunhuang GrottoMurals This paper collected the experimental data (300images) of the FlyingApsaras in theDunhuangGrottoMuralspublished in Zheng and Tairsquos [27] book These pictures coverfour historical periods the beginning stage (Northern Liang(421ndash439 AD) Northern Wei (439ndash535 AD) and WesternWei (535ndash556 AD)) the developing stage (Northern Zhou(557ndash581 AD) Sui Dynasty (581ndash618 AD) and Early Tang(618ndash712 AD)) the flourishing stage (High Tang (712ndash848AD) Late Tang (848ndash907 AD) and Five Dynasties (907ndash960 AD)) and the terminal stage (Song Dynasty (960ndash1036 AD) Western Xia (1036ndash1227 AD) and Yuan Dynasty(1227ndash1368 AD))

32 Saliency Detection Results of Four Historical Periods Allthe algorithms are used to compute the salient maps forthe same dataset Our approach is implemented in MAT-LAB R2009a software environment The other algorithmsare downloaded from the authorsrsquo homepages All thesealgorithms were implemented using Windows XP with anIntel(R) Core (TM) i3-2130 CPU 340GHz processor and4GB of memory Figure 2 shows that the salient mapsproduced by the proposed algorithm can highlight culturalelements such as streamers clouds totem and clothes ofFlying Apsaras particularly for images in the flourishingstage the line structure and spatial layout that reflect paintingstyles are highlighted for further processing that is imageenhancement or image rendering This is beneficial to theobjective of preserving the complete structural informationof salient objects and it is useful for researchers whenanalyzing the spatial structure of streamers with differentperiods

33 Comparison with Five Other State-of-the-Art ApproachesVisual saliency detection aims to produce the salient mapsof images via simulating the behavior of HVS The proposedapproaches compared with five state-of-the-art approachesdesigned for saliency detection including HC [17] LC [11]RC [17] SR [12] and IG [14] Visual saliency ground truthis constructed from eye fixation patterns of 35 observers inour experiments The state-of-the-art performance of visualsaliency detection for real scene images with fixation groundtruth similar to the case was reported by Judd et al [25]

Figure 3 illustrates that the salient area can be adequatelydistinguished from the background area via the proposedapproach One reason for this is that the contrast basedon color component and gray histogram information is


Stages Periods Original image Salient map Stages Periods Original image Salient mapTh

e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD

Figure 2 Visual saliency detection results of Flying Apsaras in the Dunhuang Grotto Murals within four historical periods the beginningstage the developing stage the flourishing stage and the terminal stage

more distinguishable than other representations the otherreason is that a spatial attention function to simulate humanrsquosvisual attention stimuli is adopted in our approach Fromperspective of highlighting painting lines of Flying Apsarasthe final salient maps by using of the proposed approach aremore consistent than the other approaches with the abilityto highlight the painting lines of Flying Apsarasrsquo streamers inthe Dunhuang Grotto Murals and guarantee their integritysimultaneously And from Figure 3 it can be concluded thatthe saliency detection results obtained by using the proposedapproach (HIG) are outperforming results obtained using thefrequency-tuned method (IG)

34 Performance Evaluation of Different Saliency DetectionApproaches In order to evaluate the pros and cons of eachapproach in a quantitative way a False Positive Rate andTrue Positive Rate (FPR-TPR) curve which is a standardtechnique for information for our Flying Apsaras datasets

retrieval community is adopted in this paper False PositiveRate and True Positive Rate (or Recall) are defined as [28ndash30]

False Positive Rate = FPFP + TN

True Positive Rate = TPTP + FN

(13)

where FP is false positive (regarding the pixels not in theobtained salient regions (background) as the pixels in theground truth) TN is true negative (regarding the pixels not inthe obtained salient regions (background) as the pixels not inthe ground truth) TP is true positive (regarding the pixels inthe obtained salient regions (foreground) as the pixels in theground truth) and FN is false negative (regarding the pixelsin the obtained salient regions (foreground) as the pixels notin the ground truth) False Positive Rate is the probability thata true negative was labeled a false positive and True PositiveRate corresponds to the fraction of the detected salient pixelsbelonging to the salient object in the ground truth


Input data SRRC IGLCHCGT HIG

Figure 3 Comparison of visual detection results via different algorithms HC [17] LC [11] RC [17] SR [12] IG [14] and HIG And all thesealgorithms are compared based on the visual saliency ground truth

Using MABLAB software the FPR-TPR curves wereplotted for the different saliency detection algorithms andshown in Figure 4 Our experiment follows the settings in[14 17] where the salient maps are digitized at each possiblethreshold within the range [0 255] Our approach achievesthe highest TPR in almost the entire FPR range [0 1] Thisis because combining saliency information from three scales

makes the background generally have low saliency valuesOnly sufficiently salient objects can be detected in this case

From these FPR-TPR curves in Figure 4 we can see thatour approach has better performances when there is lessFPR thus it can be concluded that the improved saliencydetection approach (HIG) is superior to the other five state-of-the-art approaches Achieving the sameTPR the proposed


10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate

Figure 4 FPR-TPR curves for different visual saliency detectionalgorithms HC [17] LC [11] RC [17] SR [12] IG [14] and HIG inthe experiments

algorithm has the lowest FPR which means it can detectmore salient regions and with the same FPR the proposedalgorithm has high TPR which means it can detect salientregions more accurately

In addition Precision Recall (or True Positive Rate) 119865-measure and the values of area under the FPR-TPR curves(AUC) are also used to evaluate performances of thesesaliency detection approaches quantitatively Recall is definedas above and Precision is defined as [14]

Precision = TPTP + FP

(14)

Precision corresponds to the percentage of detected salientpixels correctly assigned while Recall corresponds to thefraction of the detected salient pixels belonging to the salientobject in the ground truth High Recall can be achieved atthe expense of reducing Precision So it is necessary andimportant to measure them together therefore 119865-measure isdefined as [14]

119865-measure =(1 + 120573

2) times Precision times Recall

1205732 times Precision + Recall (15)

We use 1205732 = 03 suggested in Achanta et al [14] to weightPrecision more than Recall

In Figure 5 compared with the state-of-the-art results wecan concluded that the rank of these approaches is HIG gtHC[17] gt LC [11] gt IG [14] gt RC [17] gt SR [12] and our improvedapproach (HIG) shows the highest Precision (9013) Recall(9459) 119865-measure (9112) and AUC (9151) values instatistical sense which are a little higher than HC [17] withthe Precision (8452) Recall (9382)119865-measure (8650)and AUC (8939) values and this conclusion also can besupported by Figure 4

1

09

08

07

06

05

04

03

02

01

00SR RC IG LC HC HIG

PrecisionRecall

F-measureAUC

Figure 5 Statistical comparson results (Precision Recall 119865-measure andAUC) of different visual saliency detection algorithmsHC [17] LC [11] RC [17] SR [12] IG [14] and HIG in theexperiments

4 Conclusion and Future Works

In this paper an improved saliency detection approach basedon a frequency-tuned method for Flying Apsaras in theDunhuang Grotto Murals has been proposed This proposedapproach utilizes a CIErsquos Lab color space information andgray histogram information In addition a spatial attentionfunction to simulate humanrsquos visual attention stimuli is pro-posed which is easy to implement and provides consistentsalient maps Furthermore it is beneficial to highlight thepainting lines of Flying Apsarasrsquo streamers in the DunhuangGrotto Murals and guarantee their integrity simultaneouslyThe final experiment results on the dataset of Flying Apsarasin the Dunhuang Grotto Murals show that the proposedapproach is very effective when compared to the five state-of-the-art approaches In addition the quantitative comparisonresults demonstrated that the proposed approach outper-forms the five classical approaches in terms of FPR-TPRcurves Precision Recall 119865-measure and AUC

Future research will consider the use of deep learn-ing techniques to extract semantic information from theDunhuang Murals and based on the extracted semanticinformation visual saliency detection approaches shouldbe enhanced More recently these approaches includingDeepBeliefNetworks (DBN) RestrictedBoltzmannMachine(RBM) and Convolutional Neural Networks (CNN) couldbe applied to determine various activities in the DunhuangGrotto Murals for example the depicted musical entertain-ment scene Recognition of these scenes would be highlyvaluable in ongoing efforts to analyze Chinese cultural back-grounds and evolution in the early dynasties between 421AD and 1368 AD


Conflict of Interests

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

Acknowledgments

This research was supported in part by National Key BasicResearchProgramofChina (no 2012CB725303) theNationalScience Foundation of China (no 61170202 no 41371420 no61202287 and no 61203154) and the Fundamental ResearchFunds for the Central Universities (no 2013-YB-003) Theauthors would like to thank the anonymous reviewers and theacademic editor for the valuable comments

References

[1] F Xu L Zhang Y Zhang and W Ma ldquoSalient feature selec-tion for visual concept learningrdquo in Advances in MultimediaInformation ProcessingmdashCM 2005 vol 3767 of Lecture Notes inComputer Science pp 617ndash628 Springer BerlinGermany 2005

[2] N Imamoglu W Lin and Y Fang ldquoA saliency detection modelusing low-level features based on wavelet transformrdquo IEEETransactions on Multimedia vol 15 no 1 pp 96ndash105 2013

[3] Z Chen J Yuan and Y P Tan ldquoHybrid saliency detection forimagesrdquo IEEE Signal Processing Letters vol 20 no 1 pp 95ndash982013

[4] A Borji D N Sihite and L Itti ldquoSalient object detection abenchmarkrdquo in Proceedings of the European Conference on Com-puter Vision pp 414ndash429 2012

[5] C Koch and S Ullman ldquoShifts in selective visual attentiontowards the underlying neural circuitryrdquo in Matters of Intelli-gence pp 115ndash141 Springer AmsterdamThe Netherlands 1987

[6] C Koch and T Poggio ldquoPredicting the visual world silence isgoldenrdquo Nature Neuroscience vol 2 no 1 pp 9ndash10 1999

[7] L Itti C Koch and E Niebur ldquoAmodel of saliency-based visualattention for rapid scene analysisrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 20 no 11 pp 1254ndash12591998

[8] J Harel C Koch and P Perona ldquoGraph-based visual saliencyrdquoAdvances in Neural Information Processing Systems vol 19 pp545ndash552 2007

[9] T Liu Z Yuan J Sun et al ldquoLearning to detect a salient objectrdquoIEEE Transactions on Pattern Analysis andMachine Intelligencevol 33 no 2 pp 353ndash367 2011

[10] S Goferman L Zelnik-Manor and A Tal ldquoContext-aware sali-ency detectionrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 34 no 10 pp 1915ndash1926 2012

[11] Y Zhai and M Shah ldquoVisual attention detection in videosequences using spatiotemporal cuesrdquo in Proceedings of the14th Annual ACM International Conference on Multimedia(MULTIMEDIA rsquo06) pp 815ndash824 October 2006

[12] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Computer Society Confer-ence on Computer Vision and Pattern Recognition (CVPR rsquo07)pp 1ndash8 June 2007

[13] V Gopalakrishnan Y Hu and D Rajan ldquoRandom walks ongraphs to model saliency in imagesrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and Pattern

Recognition Workshops (CVPR rsquo09) pp 1698ndash1705 Miami FlaUSA June 2009

[14] R Achanta S Hemamiz F Estraday and S Susstrunky ldquoFre-quency-tuned salient region detectionrdquo in Proceedings of theIEEE Computer Society Conference on Computer Vision andPattern RecognitionWorkshops (CVPR rsquo09) pp 1597ndash1604 June2009

[15] Q Yan L Xu J Shi and J Jia ldquoHierarchical saliency detectionrdquoin Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo13) pp 1155ndash1162 IEEEPortland Ore USA June 2013

[16] P Siva C Russell T Xiang and L Agapito ldquoLooking beyondthe image unsupervised learning for object saliency and detec-tionrdquo in Proceedings of the 26th IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo13) pp 3238ndash3245Portland Ore USA June 2013

[17] M M Cheng G X Zhang N J Mitra X Huang and S MHu ldquoGlobal contrast based salient region detectionrdquo in Proceed-ings of the IEEE Conference on Computer Vision and PatternRecognition (CVPR rsquo11) pp 409ndash416 June 2011

[18] W M Osberger and A M Rohaly ldquoAutomatic detection ofregions of interest in complex video sequencesrdquo in 6th HumanVision and Electronic Imaging vol 4299 of Proceedings of SPIEpp 361ndash372 San Jose Calif USA June 2001

[19] C Connolly and T Fleiss ldquoA study of efficiency and accuracyin the transformation from RGB to CIE Lab color spacerdquo IEEETransactions on Image Processing vol 6 no 7 pp 1046ndash10481997

[20] R Achanta F Estrada PWils and S Susstrunk ldquoSalient regiondetection and segmentationrdquo in Proceedings of the InternationalConference on Computer Vision Systems pp 66ndash75 2008

[21] C Yang L Zhang H Lu X Ruan and M-H Yang ldquoSaliencydetection via graph-based manifold rankingrdquo in Proceedingsof the 26th IEEE Conference on Computer Vision and PatternRecognition (CVPR rsquo13) pp 3166ndash3173 June 2013

[22] D G Lowe ldquoDistinctive image features from scale-invariantkeypointsrdquo International Journal of Computer Vision vol 60 no2 pp 91ndash110 2004

[23] D Marr Vision A Computational Investigation Into the HumanRepresentation and Processing of Visual Information W HFreeman San Francisco Calif USA 1982

[24] N D B Bruce and J K Tsotsos ldquoSaliency based on informationmaximizationrdquo in Proceedings of the Annual Conference onNeural Information Processing Systems (NIPS rsquo05) pp 155ndash162December 2005

[25] T Judd K Ehinger F Durand and A Torralba ldquoLearningto predict where humans lookrdquo in Proceedings of the 12thInternational Conference on Computer Vision (ICCV rsquo09) pp2106ndash2113 IEEE Kyoto Japan October 2009

[26] H Zhang W Wang G Su and L Duan ldquoA simple andeffective saliency detection approachrdquo in Proceedings of the 21stInternational Conference on Pattern Recognition (ICPR rsquo12) pp186ndash189 November 2012

[27] R Zheng and J Tai Collected Edition of Flying Apsaras in theDunhuangGrottoMurals Commercial Press HongKong 2002(Chinese)

[28] K Bowyer C Kranenburg and S Dougherty ldquoEdge detectorevaluation using empirical ROC curvesrdquo Computer Vision andImage Understanding vol 84 no 1 pp 77ndash103 2001


[29] I E Abdou andWK Pratt ldquoQuantitative design and evaluationof enhancementthresholding edge detectorsrdquoProceedings of theIEEE vol 67 no 5 pp 753ndash763 1979

[30] D R Martin C C Fowlkes and J Malik ldquoLearning to detectnatural image boundaries using local brightness color and tex-ture cuesrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 26 no 5 pp 530ndash549 2004

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



the original state of the images in the Dunhuang GrottoMurals on the other hand the final salient map of FlyingApsaras that computed by local contrast features can extractand highlight the edge parts with striking local contrast theseparts including the painting lines have played a critical rolein analyzing significant painting styles of Flying Apsaras Inaddition visual saliency detection can be used to extractthe salient regions of Flying Apsaras and these regions canrepresent the remarkable paintingsrsquo main thought designand plot

One of the most challenging problems with perceptionis information overload This overload presents a challengewhen trying to obtain useful information from images dueto the semantic gap between low level features and high levelsemantic concepts [1] Visual attention is one of the primaryfeatures of human visual system (HVS) to derive importantand compact information from the natural scenes [2] Andits mechanism enables a reduction of the redundant data thatbenefits perception during the selective attention process Inthe image processing field saliency detection is the processof detecting the interesting visual information Furthermorethe selected stimuli need to be prioritized with the mostrelevant being processed first and the less important oneslater and it leads to a sequential treatment of different partsof the visual scenes Thus saliency detection has been widelyused in discovering regions of interests predicting humanvisual attention image quality assessment and many otherapplications [3] The saliency detection approach is used inmany fields including object detection object recognitionimage enhancement image rendering image compressionand video summarization [4]

Many different approaches to visual saliency detectionhave been proposed by a variety of researchers from dif-ferent research backgrounds As pioneers Koch Ullmanand Poggio [5 6] proposed a very influential Bioinspiredmodel and Winner-Take-All selected mechanism to simulatevisual system of humans Itti et al [7] derived bottom-up andtop-down visual saliency using center-surround differencesvia multiscale features integration Harel et al [8] proposeda new bottom-up visual saliency detection model graph-based visual saliency (GBVS) to form activation maps oncertain feature channels and normalize them with highlight-ing conspicuity and combining with other maps Liu et al[9] used the conditional random field (CRF) approach toeffectively combine a set of novel features including multi-scale contrast center-surround histogram and color spatialdistribution to describe salient object locally regionallyand globally Goferman et al [10] proposed a new type ofsaliency model context-aware saliency which is based onfour principles observed in the psychological literature locallow-level features global considerations visual organizationrules and high-level factors Zhai and Shah [11] developeda fast approach for computing pixel-level salient maps usingcolor histograms of images and a dynamic fusion techniqueapplied to combine the temporal and spatial salient mapsHou and Zhang [12] proposed a spectral residual approachfor visual saliency detection to construct the correspondingsalient maps in the spatial domain and extract the spectralresidual of an image by analyzing the log-spectrum of it

This is replaced by the phase spectrum of the Fouriertransform in Gopalakrishnan et al [13] because it is moreeffective and computationally efficient Achanta et al [14]combined low-level features of color and luminance to detectsalient regions of images which was called a frequency-tunedmethod More recently Yan et al [15] employed a multilayerapproach to analyze saliency cues and the final salient mapwas produced in a hierarchical model Siva et al [16] usedan unsupervised learning approach that annotated objects ofinterest in images to capture the salient maps

Due to employing only a few low-level features of imagesthe visual saliency detection results of the frequency-tunedmethod proposed by Achanta et al [14] are not very sat-isfactory In this paper an alternative approach histogramcombined with image average and Gaussian blur (HIG) isproposed as an improvement to the frequency-tunedmethodHIG combines features of color and luminance with thegray histogram information and uses the Euclidean distancesbetween local features and global average features to definevisual saliency as local contrast at pixel level And thefinal visual saliency definition is based on normalization ofprior visual saliency and spatial attention function Unlikeexisting approaches that focus on multiple image featuresthe proposed approach only focuses on contrast and spatialattention information from the simulation of HVS and theselocal contrast features have played a key role in highlightingthe edge parts including painting lines of Flying ApsarasThisallows for more efficient computation of salient maps andvery effective performancewhen compared to five other state-of-the-art approaches These five state-of-the-art approachesof saliency detection are Zhai and Shah [11] Hou and Zhang[12] Achanta et al [14] and Cheng et al [17] hereby referredto as LC SR IG HC and RC respectively The choice ofthese algorithms was based on their frequency of citation inliterature popularity and variety

The rest of this paper is organized as follows Section 2provides an overview of the proposed saliency detectionapproach followed by a more detailed description Section 3presents the experimental validations and comparison ofresults finally Section 4 contains the conclusion and recom-mendations for future works

2 The Proposed Saliency Detection Approach

An overview of the improved saliency detection approachproposed for Flying Apsaras in the Dunhuang Grotto Muralsin this paper is shown in Figure 1 This improved approachcomprises the following three important steps (1) imagecolor and gray channel decomposition (2) gray feature valueand color channel convolution (3) visual saliency definitionbased on normalization of prior visual saliency and spatialattention function Using these three steps the final salientmap is determined from the input image that is shown as anexample in Figure 1

21 Decomposition of Color and Gray Components fromthe Input Images It has been widely recognized amongresearchers that HVS does not equally process all


Lab image (L)

Lab image (A)

Lab image (B)

Salient map

Gauss convolution

Gre

y va

lue

Grey features value

spac

e

Col

or sp

ace

conv

ersio

n




Normaliz-

Spatialattention


I120583



|H(i j) minus

ation


rfloorrceil

I(i j) =

L(i j)

A(i j)


+

=

m

sumi=1

n

sumj=1

H(i j)(m times n)

H(i j) = 1

21205871205902eminus((iminus(m2))

2+(jminusn2))

22120590

2)

Avg LAvg AAvg B

Avg H|

Avg H

Avg H









119885 = 0010 times 119866 + 0990 times 119861

119871 = 116119891 (119884) minus 16

119886 = 500 sdot [119891 (119883

0982) minus 119891 (119884)]

119887 = 200 sdot [119891 (119884) minus 119891(119885

1183)]

(1)

where

119891 (119883) =

7787119883 + 0138 119883 le 0008856

11988313 119883 gt 0008856

(2)





Let120596119897119888and120596







119897119888and 120596



DoG (119909 119910) = 119866 (119909 119910 1205901) minus 119866 (119909 119910 120590

2)

=1

2120587[1

1205902

1

119890minus(1199092+1199102)21205902

1 minus1

1205902

2

119890minus(1199092+1199102)21205902

2]

(3)

where 1205901and 120590





11205902= 1 16 [23] If we define 120590

1= 120588120590 and 120590

2= 120590





119873minus1

sum

119899=0

[119866 (119909 119910 120588119899+1120590) minus 119866 (119909 119910 120588

119899120590)]

= 119866 (119909 119910 120588119873120590) minus 119866 (119909 119910 120590)

(4)






119878 (119909 119910) =10038171003817100381710038171003817119868120596ℎ119888

(119909 119910) minus 119868120583

10038171003817100381710038171003817 (5)

where 119868120583



2norm




2



119867(119894 119895) =1

21205871205902119890minus((119894minus1198982)

2+(119895minus1198992)

2)21205902

(6)



Avg119867 =sum119898

119894=1sum119899

119895=1119867(119894 119895)

119898 times 119899 (7)



120596ℎ119888

(119894 119895) = (119866119884119871(119894 119895) 119866119884119860(119894 119895) 119866119884119861(119894 119895))119879


120583=


Avg 119871 =sum119898

119894=1sum119899

119895=1119871 (119894 119895)

119898 times 119899 Avg119860 =

sum119898

119894=1sum119899

119895=1119860 (119894 119895)

119898 times 119899

Avg119861 =sum119898

119894=1sum119899

119895=1119861 (119894 119895)

119898 times 119899

(8)


119875(119894 119895)


119878119875(119894 119895) =

10038171003817100381710038171003817119868120596ℎ119888

(119894 119895) minus 119868120583

10038171003817100381710038171003817+1003816100381610038161003816119867 (119894 119895) minus Avg119867

1003816100381610038161003816 (9)




119878119873(119894 119895) = 255 times


(10)








119865 (119894 119895) =119871 minus 2119863 (119894 119895)

119871 + 2119863 (119894 119895) (11)



119878 (119894 119895) = 119865 (119894 119895) sdot 119878119873(119894 119895) (12)




120596ℎ119888




119873(119894 119895)












e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD







(13)









10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










Lab image (L)

Lab image (A)

Lab image (B)

Salient map

Gauss convolution

Gre

y va

lue

Grey features value

spac

e

Col

or sp

ace

conv

ersio

n




Normaliz-

Spatialattention


I120583



|H(i j) minus

ation


rfloorrceil

I(i j) =

L(i j)

A(i j)


+

=

m

sumi=1

n

sumj=1

H(i j)(m times n)

H(i j) = 1

21205871205902eminus((iminus(m2))

2+(jminusn2))

22120590

2)

Avg LAvg AAvg B

Avg H|

Avg H

Avg H









119885 = 0010 times 119866 + 0990 times 119861

119871 = 116119891 (119884) minus 16

119886 = 500 sdot [119891 (119883

0982) minus 119891 (119884)]

119887 = 200 sdot [119891 (119884) minus 119891(119885

1183)]

(1)

where

119891 (119883) =

7787119883 + 0138 119883 le 0008856

11988313 119883 gt 0008856

(2)





Let120596119897119888and120596







119897119888and 120596



DoG (119909 119910) = 119866 (119909 119910 1205901) minus 119866 (119909 119910 120590

2)

=1

2120587[1

1205902

1

119890minus(1199092+1199102)21205902

1 minus1

1205902

2

119890minus(1199092+1199102)21205902

2]

(3)

where 1205901and 120590





11205902= 1 16 [23] If we define 120590

1= 120588120590 and 120590

2= 120590





119873minus1

sum

119899=0

[119866 (119909 119910 120588119899+1120590) minus 119866 (119909 119910 120588

119899120590)]

= 119866 (119909 119910 120588119873120590) minus 119866 (119909 119910 120590)

(4)






119878 (119909 119910) =10038171003817100381710038171003817119868120596ℎ119888

(119909 119910) minus 119868120583

10038171003817100381710038171003817 (5)

where 119868120583



2norm




2



119867(119894 119895) =1

21205871205902119890minus((119894minus1198982)

2+(119895minus1198992)

2)21205902

(6)



Avg119867 =sum119898

119894=1sum119899

119895=1119867(119894 119895)

119898 times 119899 (7)



120596ℎ119888

(119894 119895) = (119866119884119871(119894 119895) 119866119884119860(119894 119895) 119866119884119861(119894 119895))119879


120583=


Avg 119871 =sum119898

119894=1sum119899

119895=1119871 (119894 119895)

119898 times 119899 Avg119860 =

sum119898

119894=1sum119899

119895=1119860 (119894 119895)

119898 times 119899

Avg119861 =sum119898

119894=1sum119899

119895=1119861 (119894 119895)

119898 times 119899

(8)


119875(119894 119895)


119878119875(119894 119895) =

10038171003817100381710038171003817119868120596ℎ119888

(119894 119895) minus 119868120583

10038171003817100381710038171003817+1003816100381610038161003816119867 (119894 119895) minus Avg119867

1003816100381610038161003816 (9)




119878119873(119894 119895) = 255 times


(10)








119865 (119894 119895) =119871 minus 2119863 (119894 119895)

119871 + 2119863 (119894 119895) (11)



119878 (119894 119895) = 119865 (119894 119895) sdot 119878119873(119894 119895) (12)




120596ℎ119888




119873(119894 119895)












e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD







(13)









10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













119885 = 0010 times 119866 + 0990 times 119861

119871 = 116119891 (119884) minus 16

119886 = 500 sdot [119891 (119883

0982) minus 119891 (119884)]

119887 = 200 sdot [119891 (119884) minus 119891(119885

1183)]

(1)

where

119891 (119883) =

7787119883 + 0138 119883 le 0008856

11988313 119883 gt 0008856

(2)





Let120596119897119888and120596







119897119888and 120596



DoG (119909 119910) = 119866 (119909 119910 1205901) minus 119866 (119909 119910 120590

2)

=1

2120587[1

1205902

1

119890minus(1199092+1199102)21205902

1 minus1

1205902

2

119890minus(1199092+1199102)21205902

2]

(3)

where 1205901and 120590





11205902= 1 16 [23] If we define 120590

1= 120588120590 and 120590

2= 120590





119873minus1

sum

119899=0

[119866 (119909 119910 120588119899+1120590) minus 119866 (119909 119910 120588

119899120590)]

= 119866 (119909 119910 120588119873120590) minus 119866 (119909 119910 120590)

(4)






119878 (119909 119910) =10038171003817100381710038171003817119868120596ℎ119888

(119909 119910) minus 119868120583

10038171003817100381710038171003817 (5)

where 119868120583



2norm




2



119867(119894 119895) =1

21205871205902119890minus((119894minus1198982)

2+(119895minus1198992)

2)21205902

(6)



Avg119867 =sum119898

119894=1sum119899

119895=1119867(119894 119895)

119898 times 119899 (7)



120596ℎ119888

(119894 119895) = (119866119884119871(119894 119895) 119866119884119860(119894 119895) 119866119884119861(119894 119895))119879


120583=


Avg 119871 =sum119898

119894=1sum119899

119895=1119871 (119894 119895)

119898 times 119899 Avg119860 =

sum119898

119894=1sum119899

119895=1119860 (119894 119895)

119898 times 119899

Avg119861 =sum119898

119894=1sum119899

119895=1119861 (119894 119895)

119898 times 119899

(8)


119875(119894 119895)


119878119875(119894 119895) =

10038171003817100381710038171003817119868120596ℎ119888

(119894 119895) minus 119868120583

10038171003817100381710038171003817+1003816100381610038161003816119867 (119894 119895) minus Avg119867

1003816100381610038161003816 (9)




119878119873(119894 119895) = 255 times


(10)








119865 (119894 119895) =119871 minus 2119863 (119894 119895)

119871 + 2119863 (119894 119895) (11)



119878 (119894 119895) = 119865 (119894 119895) sdot 119878119873(119894 119895) (12)




120596ℎ119888




119873(119894 119895)












e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD







(13)









10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











119878 (119909 119910) =10038171003817100381710038171003817119868120596ℎ119888

(119909 119910) minus 119868120583

10038171003817100381710038171003817 (5)

where 119868120583



2norm




2



119867(119894 119895) =1

21205871205902119890minus((119894minus1198982)

2+(119895minus1198992)

2)21205902

(6)



Avg119867 =sum119898

119894=1sum119899

119895=1119867(119894 119895)

119898 times 119899 (7)



120596ℎ119888

(119894 119895) = (119866119884119871(119894 119895) 119866119884119860(119894 119895) 119866119884119861(119894 119895))119879


120583=


Avg 119871 =sum119898

119894=1sum119899

119895=1119871 (119894 119895)

119898 times 119899 Avg119860 =

sum119898

119894=1sum119899

119895=1119860 (119894 119895)

119898 times 119899

Avg119861 =sum119898

119894=1sum119899

119895=1119861 (119894 119895)

119898 times 119899

(8)


119875(119894 119895)


119878119875(119894 119895) =

10038171003817100381710038171003817119868120596ℎ119888

(119894 119895) minus 119868120583

10038171003817100381710038171003817+1003816100381610038161003816119867 (119894 119895) minus Avg119867

1003816100381610038161003816 (9)




119878119873(119894 119895) = 255 times


(10)








119865 (119894 119895) =119871 minus 2119863 (119894 119895)

119871 + 2119863 (119894 119895) (11)



119878 (119894 119895) = 119865 (119894 119895) sdot 119878119873(119894 119895) (12)




120596ℎ119888




119873(119894 119895)












e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD







(13)









10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











119865 (119894 119895) =119871 minus 2119863 (119894 119895)

119871 + 2119863 (119894 119895) (11)



119878 (119894 119895) = 119865 (119894 119895) sdot 119878119873(119894 119895) (12)




120596ℎ119888




119873(119894 119895)












e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD







(13)









10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











e beg

inni

ng st

age

421

ndash556

AD

Wes

tern

Wei

535

ndash556

AD

N

orth

ern

Wei

439

ndash535

AD

N

orth

ern

Lian

g421

-439

AD

The fl

ouris

hing

stag

e712

ndash960

AD

Five

Dyn

astie

s907

ndash960

AD

La

te T

ang

848

ndash907

AD

H

igh

Tang

712

ndash848

AD

The d

evelo

ping

stag

e557

ndash712

AD

Nor

ther

n Zh

ou557

ndash581

AD

Su

i Dyn

asty

581

ndash618

AD

Ea

rly T

ang

618

ndash712

AD

960

ndash1368

AD

Th

e ter

min

al st

age

Song

Dyn

asty

960

ndash1036

AD

W

este

rn X

ia1036

ndash1227

AD

Yu

an D

ynas

ty1227

ndash1368

AD







(13)









10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















10908070605040302010

1

09

08

07

06

05

04

03

02

01

0

SR RCIGLC HC

HIG

True

pos

itive

rate

False positive rate





(14)


119865-measure =(1 + 120573





1

09

08

07

06

05

04

03

02

01


PrecisionRecall

F-measureAUC








Acknowledgments


References



































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












Acknowledgments


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 an improved saliency detection approach ...apsaras with di erent historical periods...

Documents