color image segmentation based on different color space models using automatic grabcut

Research ArticleColor Image Segmentation Based on DifferentColor Space Models Using Automatic GrabCut

Dina Khattab1 Hala Mousher Ebied1 Ashraf Saad Hussein2 and Mohamed Fahmy Tolba1

1 Faculty of Computer and Information Sciences Ain Shams University Cairo 11566 Egypt2 Faculty of Computer Studies Arab Open University-Headquarters 13033 Al-Safat Kuwait

Correspondence should be addressed to Dina Khattab dinaredakhattabgmailcom

Received 1 June 2014 Revised 16 August 2014 Accepted 16 August 2014 Published 31 August 2014

Academic Editor Shifei Ding

Copyright copy 2014 Dina Khattab 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 comparative study using different color spaces to evaluate the performance of color image segmentationusing the automatic GrabCut technique GrabCut is considered as one of the semiautomatic image segmentation techniques sinceit requires user interaction for the initialization of the segmentation processThe automation of the GrabCut technique is proposedas a modification of the original semiautomatic one in order to eliminate the user interaction The automatic GrabCut utilizesthe unsupervised Orchard and Bouman clustering technique for the initialization phase Comparisons with the original GrabCutshow the efficiency of the proposed automatic technique in terms of segmentation quality and accuracy As no explicit color spaceis recommended for every segmentation problem automatic GrabCut is applied with 119877119866119861 119867119878119881 119862119872119884 119883119884119885 and 119884119880119881 colorspaces The comparative study and experimental results using different color images show that 119877119866119861 color space is the best colorspace representation for the set of the images used

1 Introduction

The process of partitioning a digital image into multiplesegments is defined as image segmentation Segmentationaims to divide an image into regions that can be more repre-sentative and easier to analyze Such regions may correspondto individual surfaces objects or natural parts of objectsTypically image segmentation is the process used to locateobjects and boundaries (eg lines or curves) in images [1]Furthermore it can be defined as the process of labeling everypixel in an image where all pixels having the same labelshare certain visual characteristics [2] Usually segmentationuses local information in the digital image to compute thebest segmentation such as color information used to createhistograms or information indicating edges boundaries ortexture information [3]

Color image segmentation that is based on the colorfeature of image pixels assumes that homogeneous colors inthe image correspond to separate clusters and hence mean-ingful objects in the image In other words each clusterdefines a class of pixels that share similar color properties

As the segmentation results depend on the used color spacethere is no single color space that can provide acceptableresults for all kinds of images For this reason many authorstried to determine the color space that will suit their specificcolor image segmentation problem [4] In this work a seg-mentation of color images is tested with different classicalcolor spaces 119877119866119861 119862119872119884119883119884119885 119884119880119881 and119867119878119881 to select thebest color space for the considered kind of images

The segmentation process is based on the GrabCutsegmentation technique [5] which is considered as one ofthe powerful state-of-the-art techniques for the problem ofcolor image segmentationThe iterative energy minimizationscheme of the GrabCut is based on the powerful optimiza-tion of the Graph Cut technique [6] which allows for thegeneration of the global optimal segmentation In additionGraph Cut can be easily well extended to the problem of N-Dimages Furthermore the cost energy function of the GraphCut minimization process allows it to be defined in terms ofdifferent image features such as color region boundary orany mixture of image features This flexibility provides widepotential for the use of GrabCut in different applications

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 126025 10 pageshttpdxdoiorg1011552014126025

2 The Scientific World Journal

On the other hand GrabCut is considered as a bilabelsegmentation technique where images can be segmentedinto two background and foreground regions only Initialuser intervention is required in order to specify an objectof interest to be segmented out of the image considering allthe remaining image pixels as one background region Thisclassifies the GrabCut as a semiautomatic segmentation tech-nique and turns the quality of the initialization and hencethe segmentation performance sensitive to the user selectionIn other words poor GrabCut initialization may lead to badfinal segmentation accuracy which might require extra userinteractions with the segmentation results for fine tuning [5]

In this work a modified GrabCut is proposed as an auto-matic segmentation technique which can segment the imageinto its natural objects without any need for the initialuser intervention Automation of GrabCut is applied usingOrchard and Bouman clustering [7] as an unsupervised clus-tering technique The selection of the Orchard and Boumanclustering is based on the empirical comparison results car-ried out in the work of [8] The paper exploits the use ofsome evaluation criteria to evaluate the discriminating powerof the automatic GrabCut with the different color spacesThe remainder of the paper is organized as follows Section 2provides a basic background on segmentation based-colorspacemodels image segmentation using GrabCut and unsu-pervised clustering techniques Section 3 explains the dif-ferent color space models Section 4 illustrates the Orchardand Bouman clustering The original GrabCut techniqueand details of its modification are explained in Section 5Experimental results are presented in Section 6 while theconclusion and future work are presented in Section 7

2 Related Work

As no common opinion has emerged about which is thebest choice for color space based image segmentation someresearch work tried to identify the best color space for aspecific task Several works [9 10] show that different colorspaces are useful for the problem of color image segmenta-tion Jurio et al [11] have carried out a comparative studybetweendifferent color spaces in cluster based image segmen-tation using two similar clustering algorithms Their studyinvolved the test of four color spaces 119877119866119861 119867119878119881 119862119872119884and 119884119880119881 in order to identify the best color representationThey obtained their best results in most cases using 119862119872119884

color space while 119867119878119881 also provided good results Businet al [4] proposed a method to automatically select a specificcolor space among classical color spaces This selection wasdone according to an evaluation criterion based on a spectralcolor analysis This criterion evaluates the quality of thesegmentation in each space and selects the best one whichpreserves its own specific properties A study of the ten mostcommon color spaces for skin color detection was presentedin [12] They concluded that 119867119878119881 is the best color spaceto detect skin in an image Another study that was appliedfor the classification of pizza toppings [13] proved that thepolynomial SVM classifier combined with 119867119878119881 color spaceis the best approach among five different color spaces Basedon a comparative study between the 119877119866119861 and 119867119878119881 models

Ruiz-Ruiz et al [14] declared that the best accuracy wasachieved with 119867119878119881 representation in order to achieve realtime processing in real farm fields for crop segmentation

GrabCut is considered one of the powerful techniquesused for color image segmentation It has been appliedto different segmentation problems such as human bodysegmentation [15ndash17] video segmentation [18] semanticsegmentation [19] and volume segmentation [20] In [17]an automatic extraction of the human body from colorimages was developed byHuThe iteratedGrabCut techniquewas used to dynamically update a trimap contour whichwas initialized from the results of a scanning detector usedfor detecting faces from images The research has somedrawbacks as the process goes through many steps anditerations in addition to being constrained to human poseswith frontal side faces A fully automatic Spatio-TemporalGrabCut human segmentation methodology was proposedby Hernandez et al [16] They developed methodologythat takes the benefits of the combination of tracking andsegmentation Instead of the initial user intervention toinitialize theGrabCut algorithm a set of seeds defined by facedetection and a skin color model are used for initializationAnother approach to segment humans from cluttered imageswas proposed by Gulshan et al in [15] They utilized thelocal colormodel basedGrabCut for automatic segmentationThis GrabCut local color model was used to refine the crudehuman segmentations they obtained In video segmentationCorrigan et al [18] extended GrabCut for more robust videoobject segmentation They extended the Gaussian mixturemodel (GMM) of the GrabCut algorithm so that the colorspace was complemented with the derivative in time of thepixelrsquos intensities in order to include temporal informationin the segmentation optimization process Goring et al [19]integrated GrabCut into a semantic segmentation frameworkby labeling objects in a given image Most recently Ramırezet al [20] proposed a fully parallelized scheme usingGrabCutfor 3D segmentation that has been adopted to run on GPUThe scheme aims at producing efficient segmentation resultsfor the case of volume meshes in addition to reducing thecomputational time

Clustering [21] the unsupervised classification of pat-terns into groups is one of the most important tasks inexploratory data analysis [22] It has a long and rich historyin a variety of scientific disciplines including anthropologybiology medicine psychology statistics mathematics engi-neering and computer science Clustering in image segmen-tations [2 23 24] is defined as the process of identifyinggroups of similar image primitives Unsupervised clusteringtechniques [25] are content based clustering where contentrefers to shapes textures or any other information that canbe inherited from the image itself

In the cases of bilabel segmentation good separationbetween foreground and background is requiredThis can beimplemented through finding clusters with a low variancesince this makes the cluster easier to separate from the othersThe selection of the Orchard and Bouman clustering tech-nique [7] is guided by Ruzon and Tomasi [26] and Chaunget al [27] in order to get tight and well separated clustersThey have worked on solving the problem of image matting

The Scientific World Journal 3

that is required for image compositing In their approachOrchard and Bouman binary split algorithm has been usedfor partitioning the unknown region colors into several clus-ters in order to generate a color distribution for the unknownregion to be estimated According to a comparative study in[8] the Orchard and Bouman clustering outperformed otherunsupervised clustering techniques including self-organizingmaps (SOFM) and fuzzy C-means (FCM) for the automationof the GrabCut in terms of improving the segmentationaccuracy

3 Color Space Models

The most widely used color space is the 119877119866119861 color spacewhere a color point in the space is characterized by threecolor components of the corresponding pixel which are red(119877) green (119866) and blue (119861) However since there exist alot of color spaces it is useful to classify them into fewercategories with respect to their definitions and propertiesVandenbroucke [28] proposed the classification of the colorspaces into the following categories

(i) The primary spaces which are based on the theorythat assumes it is possible to match any color bymixing an appropriate amount of the three primarycolors the primary spaces are the real 119877119866119861 thesubtractive 119862119872119884 and the imaginary 119883119884119885 primaryspaces The conversion from 119877119866119861 to 119862119872119884 is

1198621015840= 1 minus 119877 119862 = min (1max (0 1198621015840 minus 119870

1015840))

1198721015840= 1 minus 119866 119872 = min (1max (01198721015840 minus 119870

1015840))


1015840))

1198701015840= min (11986210158401198721015840 1198841015840)

(1)

and the conversion from 119877119866119861 to119883119884119885 is

[

[

119883

119884

119885

]

]

= [

[

0412453 0357580 0180423

0212671 0715160 0072169

0019334 0119193 0950227

]

]

[

[

119877

119866

119861

]

]

(2)

(i) The luminance-chrominance spaces which are com-puted of one color component that represents theluminance and two color components that representthe chrominance the 119884119880119881 color space is an exampleof the luminance-chrominance spaces The conver-sion from 119877119866119861 to 119884119880119881 is

[

[

119884

119880

119881

]

]

= [

[

02989 05866 01145

minus0147 minus0289 0436

0615 minus0515 minus0100

]

]

[

[

119877

119866

119861

]

]

(3)

(ii) The perceptual spaces that try to quantify the sub-jective human color perception by means of threemeasures intensity hue and saturation the 119867119878119881

is an example of the perceptual color space Theconversion from 119877119866119861 to119867119878119881 is

119867 =

0 if Max = Min

(60∘times

119866 minus 119861

MaxminusMin+ 360

∘)

times mod 360∘ if Max = 119877

60∘times

119861 minus 119877

MaxminusMin+ 120

∘ if Max = 119866

60∘times

119877 minus 119866

MaxminusMin+ 240

∘ if Max = 119861

119878 =

0 if max = 0

MaxminusMinMax

otherwise

119881 = Max

(4)

4 Orchard and Bouman Clustering Technique

Orchard and Bouman [7] is a color quantization clusteringtechnique that uses the eigenvector of the color covariancematrix to determine good cluster splits The algorithm startswith all the pixels in a single cluster The cluster is then splitinto two using a function of eigenvector of the covariancematrix as the split point Then it uses the eigenvalues of thecovariance matrices to choose which of the resulting clustersis candidate for the next splitting This procedure is repeateduntil the desired number of clusters is achieved It is an opti-mal solution for large clusters with Gaussian distributions

For example consider 1198621as a set of pixels in order to

divide it into 119870 clusters

(1) calculate 1205831 the mean of 119862

1 and Σ

1 the covariance

matrix of 1198621

(2) for 119894 = 2 to 119870 do the following

(i) find the set 119862119899which has the largest eigenvalue

and store the associated eigenvector 119890119899

(ii) split 119862119899into two sets 119862

119894= 119909 isin 119862

119899 119890119879

119899119911119899le

119890119879

119899120583119899 and 119862lowast

119899= 119862119899minus 119862119894

(iii) compute 120583lowast119899 Σlowast119899 120583119894 and Σ

119894

This results in119870 pixel clusters

5 Image Segmentation Using GrabCut

Image segmentation is simply the process of separating animage into foreground and background parts Graph Cuttechnique [6] was considered as an effective way for thesegmentation of monochrome images which is based on theMin-CutMax-Flow algorithm [29] GrabCut [5] is a power-ful extension of the Graph Cut algorithm to segment colorimages iteratively and to simplify the user interaction neededfor a given quality of the segmentation results Section 51


(a) (b)

Figure 1 Example of GrabCut segmentation (a) GrabCut allows the user to drag a rectangle around the object of interest to be segmented(b) The segmented object

explains the original semiautomatic GrabCut algorithm asdeveloped by Rother et al in [5] while its modification forautomatic segmentation is presented in Section 52

51 Original Semiautomatic GrabCut The GrabCut algo-rithm learns the color distributions of the foreground andbackground by giving each pixel a probability to belong toa cluster of other pixels It can be explained as follows givena color image I let us consider the 119911 = (119911

1 119911

119899 119911

119873)

of 119873 pixels where 119911119894= (1198621119894 1198622119894 1198623119894) 119894 isin [1 119873] and

119862119895is the 119895th color component in the used color space The

segmentation is defined as an array 120572 = (1205721 120572

119873) 120572119894isin

0 1 assigning a label to each pixel of the image indicating ifit belongs to the background or the foregroundThe GrabCutalgorithm consists mainly of two basic steps initializationand iterative minimization The details of both steps areexplained in the following subsections

511 GrabCut Initialization The novelty of the GrabCuttechnique is in the ldquoincomplete labelingrdquo which allows areduced degree of user interaction The user interactionconsists simply of specifying only the background pixels bydragging a rectangle around the desired foreground object(Figure 1) The process of GrabCut initialization works asfollows

Step 1 A trimap 119879 = TBTUTF is initialized in asemiautomatic way The two regions TB and TU contain theinitial background and uncertain pixels respectively whileTF = OslashThe initial TB is determined as the pixels around theoutside of the marked rectangle Pixels belonging to TB areconsidered as a fixed background whereas those belongingto TU will be labeled by the algorithm

Step 2 An initial image segmentation 120572 = (1205721 120572

119894

120572119873) 120572119894

isin 0 1 is created where all unknown pixelsare tentatively placed in the foreground class (120572

119894= 1 for

119894 isin TU) and all known background pixels are placed in thebackground class (120572

119894= 0 for 119894 isin TB)

Step 3 Two full covarianceGaussianmixturemodels (GMMs)are defined each consisting of 119870 = 5 components one for

background pixels (120572119894= 0) and the other one for foreground

(initially unknown) pixels (120572119894= 1) The 119870 components of

both GMMs are initialized from the foreground and back-ground classes using the Orchard and Bouman clusteringtechnique

512 GrabCut Iterative Energy Minimization The final seg-mentation is performed using the iterative minimizationalgorithm of the Graph Cut [6] in the following steps

Step 4 Each pixel in the foreground class is assigned to themost likely Gaussian component in the foreground GMMSimilarly each pixel in the background is assigned to themostlikely background Gaussian component

Step 5 The GMMs are thrown away and new GMMs arelearned from the pixel sets created in the previous set

Step 6 A graph is built and Graph Cut is run to find a newforeground and background classification of pixels

Step 7 Steps 4ndash6 are repeated until the classification con-verges

This has the advantage of allowing the automatic refine-ment of the opacities 120572 as newly labeled pixels from the TUregion of the initial trimap are used to refine the color of theGMM

52 Proposed Automatic GrabCut Although the incompleteuser labeling of GrabCut reduces the user interaction sub-stantially it is still a requirement in order to initiate thesegmentation processThis identifies GrabCut as a semiauto-maticsupervised segmentation algorithm In order to allowthe image to be segmented into proper segments withoutany user guidance this requires replacing the semiauto-maticsupervised step of GrabCut initialization with a totallyautomaticunsupervised one

In this paper the Orchard and Bouman [7] is proposed tobe used as an image clustering technique to automatically setthe initial trimap 119879 and the initial segmentation (Section 51Steps 1 and 2) The distinction between the trimap and


1 2 3 4 5 6 7

8 9 10 11 12 13 14

15 16 17 18 19 20 21

22 23

Figure 2 The dataset of images

the segmentation formalizes the separation between theregion of interest to be segmented and the final segmentationderived by the GrabCut algorithm In the automatic tech-nique Steps 1 and 2 of the GrabCut initialization process willbe modified as follows

Step 1 While the original GrabCut constructs a trimap 119879 oftwo regions TB and TU as a fixed background and unknownregions respectively the proposed automatic technique con-siders the whole image as one unknown region TU whereTU = 119911

119894isin 1199111 119911

119899 119911

119873 119894 isin [1 119873] This means

that no fixed foreground or background regions are knownand all image pixels will be involved in the minimizationprocess to be labeled by the algorithm

Step 2 The image is initially separated into two foregroundTF and background TB regions using the Orchard andBouman clustering technique During this step a new GMMis introduced which consists of only two components (119870 =

2) one component for the background pixels (120572119894= 0) and

the other for the foreground pixels (120572119894= 1)The Orchard and

Bouman clustering technique is then applied and repeateduntil reaching the number of components (119870 = 2) in theGMM resulting in separating the image exactly into twoclusters

Step 3 The colors of image pixels belonging to each clus-ter (foreground and background clusters) generated from

the previous step are then used to initialize another two fullcovariance Gaussian mixture models (GMMs) with (119870 = 5)

Steps 4ndash7 The learning portion of the algorithm runs exactlyas the original GrabCut (Section 51 Steps 4ndash7)

6 Results and Discussions

The automatic GrabCut technique was experimentally testedusing a dataset of 23 different images as shown in Figure 2According to literature many recent works in the fieldsof cluster based image segmentation and automatic imagesegmentation are conducting their experiments on fewernumbers of images such as [9 11 30 31] They are using adataset of 8 4 4 and 15 images respectively In this workusing a dataset of 23 images can be considered a reasonablenumber of test cases This dataset is collected partially fromthe Berkeley segmentation dataset [32] and from publicallyavailable images [33] in a way that matches certain criteriaThese criteria consider a special fitting into two class seg-mentations including having mainly one object (as a fore-ground) and a well separation between the foreground andbackground color regions

For evaluation it was noticed that no binary segmen-tations exist as part of the human segmentations includedin the Berkeley segmentation dataset [32] For this reasonthe ground truth data for our selected dataset is manually


(a) (b) (c)

Figure 3 Samples of the manual binary ground truths generated

(a)

(b)

Figure 4 Visual comparison of the segmentation results of (a) original semiautomatic GrabCut and (b) automatic GrabCut initialized usingOrchard and Bouman

generated using standard image processing tools (AdobePhotoshop) Figure 3 displays samples of the manual binaryground truths generatedThe error rate and the overlap scorerate are used as two evaluation metrics The error rate iscalculated as the fraction of pixels with wrong segmentations(compared to ground truth) divided by the total number ofpixels in the image The overlap score rate is given by 119910

1cap

11991021199101cup1199102 where119910

1and1199102are any two binary segmentations

In the first experiment automatic GrabCut which is ini-tialized usingOrchard and Bouman is applied and comparedto the original GrabCut algorithm Figure 4 shows samplevisual results for the segmentation using (119870 = 5) componentsfor GMMs as recommended by Rother et al [5] Table 1shows the quantitative comparison between the original andmodified GrabCut for the whole dataset as presented inFigure 2 As shown in Table 1 the automatic GrabCut usingOrchard and Bouman clustering outperforms the originalone in terms of minimizing the error and improving thesegmentation accuracy The average error rate is 364 forthe automatic GrabCut compared to 428 for the original

GrabCut technique The overall performance looks better interms of the standard deviation (SD) which exhibits 361for the automatic GrabCut compared to 55 for the originalGrabCut

Some cases with bad segmentation error using the orig-inal GrabCut can be noticed in Table 1 (images 1 and 9)This explains one main drawback of the original GrabCutinitialization whichmakes the segmentation results sensitiveto the user selection of the area of interest to be segmentedThis occurs when other objects which are out of interestmay be considered as part of the foreground by being locatedwithin the area of the dragged rectangular boundary aroundthe object of interest The segmentation results of thesetwo images are visually illustrated in Figure 4(a) It canbe noticed how a large portion of the leaf appears in thefinal segmentation of the insect image The same problemoccurred when considering the land as part of the foregroundarea with the elephant image The quantitative comparisonsof the error rates generated for these two images in Table 1and visual comparisons in Figure 4 illustrate the efficiency


Table 1 Comparisons between the original and automatic GrabCut

Image Error rate Overlap score rate

Original semiautomatic GrabCut Automatic GrabCut usingOrchard and Bouman Original semiautomatic GrabCut Automatic GrabCut using

Orchard and Bouman1 305 1870 9591 58572 1548 574 4396 75693 479 731 9151 85484 307 309 9706 97025 416 375 8242 85186 086 086 9717 97177 240 240 6900 69018 087 108 9276 90819 2581 217 6766 973110 282 281 9632 963511 205 205 9704 970412 499 493 8845 893213 228 230 9571 956414 236 256 9685 964115 278 350 9414 919816 316 302 9338 938017 208 210 9519 951118 388 386 9095 910619 288 292 9344 933020 143 144 9647 964321 127 127 9462 946422 330 316 9337 937323 257 258 9375 9374Avg 428 364 8944 9021SD 550 361 1276 987

of the automatic GrabCut in handling such a problem Theefficiency of the automatic GrabCut is provoked by prevent-ing any hard constraints to be specified during initializationeither for foreground or background (Section 52 Step 1)

In the second experiment the automatic GrabCut whichis initialized using Orchard and Bouman is applied withdifferent color space models including 119877119866119861 119883119884119885 119862119872119884119884119880119881 and 119867119878119881 The features that identify each image pixelare only the values of its three components in the selectedcolor space The final segmentation results are obtained forall used images For a quantitative comparison Table 2 showsthe error rate and the overlap score rate for the whole datasetThe results in Table 2 are ordered in ascending order fromleft to right in terms of the total number of good imagesegmentation results and the average error rates We can seethat the 119877119866119861 space is the one that obtains better results formost of the images in terms of the average error rate 119884119880119881and 119883119884119885 follow with very little increase in the average errorrate They exhibit almost the same average error and overlapscore rates which are 549 for the error rate and 9535 forthe overlap score rate and 563 for the error rate and 9579for the overlap score rate respectively Figure 5 shows visual

segmentation results for some images while Figure 6 showsgraph plots of the average segmentation error rate and theoverlap score rate for all different color spaces

7 Conclusions and Future Work

In this paper a modification of GrabCut is presented toeliminate the need of initial user interaction for guidingsegmentation and hence converting GrabCut into an auto-matic segmentation technique The modification includesusing Orchard and Bouman as an unsupervised clusteringtechnique to initialize the GrabCut segmentation processBased on a dataset of 23 images the experiments revealedthat automatic GrabCut using Orchard and Bouman cluster-ing outperforms the original GrabCut It reduces the needfor user intervention while segmentation and adds extraadvantage for the GrabCut via automation Furthermore itprovides robust and accurate segmentationwith average errorrates of 364 compared to the results of 428 average errorrate that is achieved by the original GrabCut In additionthe performance of the automatic GrabCut is evaluated usingfive different color spaces 119877119866119861 119884119880119881 119883119884119885 119867119878119881 and


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

Figure 5 Visual comparison of segmentation results for automatic GrabCut applied in the (a) 119877119866119861 (b) 119884119880119881 (c) 119883119884119885 (d) 119862119872119884 and (e)119867119878119881 color spaces


Table 2 Experimental segmentation results on different color spaces using automatic GrabCut

Image Error rate Overlap score rate RGB YUV XYZ CMY HSV RGB YUV XYZ CMY HSV

1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()

Error Overlap scoreAverage accuracy measures

120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV

Figure 6 Comparison of average accuracy measures for applyingautomatic GrabCut segmentation on different color spaces

119862119872119884 The experimental results show that the segmentationresults depending on the 119877119866119861 color space provided the bestsegmentation results compared to other color spaces for theconsidered set of images

This study can be improved by enlarging the dataset andincluding different kinds of images On the other hand futurework might include modifying the energy minimizationprocedure of the automatic GrabCut to allow for multilabeloptimization and segmentation

Conflict of Interests

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

References

[1] R C Gonzalez and R E Woods Digital Image ProcessingPrentice-Hall 3rd edition 2006

[2] M Lalitha M Kiruthiga and C Loganathan ldquoA survey onimage segmentation through clustering algorithmrdquo Interna-tional Journal of Science and Research vol 2 no 2 pp 348ndash3582013

[3] N Sharma M Mishra and M Shrivastava ldquoColour image seg-mentaion techniques and issues an approachrdquo International Jour-nal of ScientificampTechnologyResearch vol 1 no 4 pp 9ndash12 2012

[4] L Busin N Vandenbroucke and L Macaire ldquoColor spaces andimage segmentationrdquoAdvances in Imaging and Electron Physicsvol 151 pp 65ndash168 2008


[5] Rother C V Kolmogorov and A Blake ldquoldquoGrabCutrdquo inter-active foreground extraction using iterated graph cutsrdquo ACMTransactions on Graphics vol 23 no 3 pp 309ndash314 2004

[6] Y Y Boykov andM-P Jolly ldquoInteractive graph cuts for optimalboundary amp region segmentation of objects in N-D imagesrdquoin Proceedings of the 8th International Conference on ComputerVision (ICCV rsquo01) vol 1 pp 105ndash112 IEEE Vancouver CanadaJuly 2001

[7] M T Orchard and C A Bouman ldquoColor quantization ofimagesrdquo IEEE Transactions on Signal Processing vol 39 no 12pp 2677ndash2690 1991

[8] D Khattab H M Ebied A S Hussien and M F Tolba ldquoAuto-matic GrabCut based on unsupervised clustering for imagesegmentationrdquo Journal of Computer Science and TechnologySubmitted

[9] O Alata and L Quintard ldquoIs there a best color space for colorimage characterization or representation based on multivariategaussian mixture modelrdquo Computer Vision and Image Under-standing vol 113 no 8 pp 867ndash877 2009

[10] M Pagola R Ortiz I Irigoyen et al ldquoNew method to assessbarley nitrogen nutrition status based on image colour analysiscomparison with SPAD-502rdquo Computers and Electronics inAgriculture vol 65 no 2 pp 213ndash218 2009

[11] A Jurio M Pagola M Galar C Lopez-Molina and D Pater-nain ldquoA comparison study of different color spaces in clusteringbased image segmentationrdquo in Information Processing andMan-agement of Uncertainty in Knowledge-Based Systems Applica-tions vol 81 of Communications in Computer and InformationScience pp 532ndash541 Springer 2010

[12] J M Chaves-Gonzalez M A Vega-Rodrıguez J A Gomez-Pulido and J M Sanchez-Perez ldquoDetecting skin in face recog-nition systems a colour spaces studyrdquo Digital Signal Processingvol 20 no 3 pp 806ndash823 2010

[13] C-J Du and D-W Sun ldquoComparison of three methods forclassification of pizza topping using different colour space trans-formationsrdquo Journal of Food Engineering vol 68 no 3 pp 277ndash287 2005

[14] G Ruiz-Ruiz J Gomez-Gil and L M Navas-Gracia ldquoTestingdifferent color spaces based on hue for the environmentallyadaptive segmentation algorithm (EASA)rdquoComputers and Elec-tronics in Agriculture vol 68 no 1 pp 88ndash96 2009

[15] V Gulshan V Lempitsky and A Zisserman ldquoHumanisingGrabCut Learning to segment humans using the Kinectrdquoin Proceedings of the IEEE International Conference on Com-puter Vision Workshops (ICCV Workshops rsquo11) pp 1127ndash1133Barcelona Spain November 2011

[16] A Hernandez M Reyes S Escalera and P Radeva ldquoSpatio-Temporal GrabCutt human segmentation for face and poserecoveryrdquo in Proceedings of the IEEE Computer Society Confer-ence on Computer Vision and Pattern Recognition Workshops(CVPRW rsquo10) San Francisco Calif USA June 2010

[17] Y HuHuman Body Region Extraction from Photos MVA 2007[18] D Corrigan S Robinson and A Kokaram ldquoVideo matting

using motion extended GrabCutrdquo in Proceedings of the IETEuropean Conference on Visual Media Production (CVMP rsquo08)London UK 2008

[19] C Goring B Frohlich and J Denzler ldquoSemantic segmentationusing GrabCutrdquo in Proceedings of the International Conferenceon Computer Vision Theory and Applications (VISAPP rsquo12) pp597ndash602 February 2012

[20] J Ramırez P Temoche and R Carmona ldquoA volume segmenta-tion approach based on GrabCutrdquo CLEI Electronic Journal vol16 no 2 2013

[21] R Kaur and G S Bhathal ldquoA survey of clustering techniquesrdquoInternational Journal of Advanced Research in Computer Scienceand Software Engineering vol 3 no 5 pp 153ndash157 2013

[22] A K Jain M N Murty and P J Flynn ldquoData clustering areviewrdquo ACM Computing Surveys vol 31 no 3 pp 316ndash3231999

[23] A Gulhane P L Paikrao and D S Chaudhari ldquoA review ofimage data clustering techniquesrdquo International Journal of SoftComputing and Engineering vol 2 no 1 pp 212ndash215 2012

[24] S Naz H Majeed and H Irshad ldquoImage segmentation usingfuzzy clustering a surveyrdquo inProceedings of the 6th InternationalConference on Emerging Technologies (ICET rsquo10) pp 181ndash186October 2010

[25] N Grira M Crucianu and N Boujemaa ldquoUnsupervised andsemisupervised clustering a brief surveyrdquo in Proceedings of the7th ACM SIGMM International Workshop on Multimedia Infor-mation Retrieval 2005

[26] M A Ruzon and C Tomasi ldquoAlpha estimation in naturalimagesrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition pp 18ndash25 IEEE 2000

[27] Y-Y Chuang B Curless D H Salesin and R Szeliski ldquoABayesian approach to digitalmattingrdquo in Proceedings of the IEEEComputer Society Conference on Computer Vision and PatternRecognition (CVPR rsquo01) pp II264ndashII271 December 2001

[28] N Vandenbroucke L Macaire and J-G Postaire ldquoColor imagesegmentation by pixel classification in an adapted hybrid colorspace Application to soccer image analysisrdquo Computer Visionand Image Understanding vol 90 no 2 pp 190ndash216 2003

[29] Y Boykov and V Kolmogorov ldquoAn experimental comparisonof min-cutmax-flow algorithms for energy minimization invisionrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 26 no 9 pp 1124ndash1137 2004

[30] H S Kumar K Raja K Venugopal and L Patnaik ldquoAutomaticimage segmentation using waveletsrdquo International Journal ofComputer Science and Network Security vol 9 no 2 pp 305ndash313 2009

[31] C V Narayana E S Reddy andM S Prasad ldquoAutomatic imagesegmentation using ultra fuzzinessrdquo International Journal ofComputer Applications vol 49 pp 6ndash13 2012

[32] D Martin C Fowlkes D Tal and J Malik ldquoA database ofhuman segmented natural images and its application to evalu-ating segmentation algorithms andmeasuring ecological statis-ticsrdquo in Proceedings of the 8th IEEE International Conference onComputer Vision (ICCV rsquo01) 2001

[33] httpwwwgooglecom

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



On the other hand GrabCut is considered as a bilabelsegmentation technique where images can be segmentedinto two background and foreground regions only Initialuser intervention is required in order to specify an objectof interest to be segmented out of the image considering allthe remaining image pixels as one background region Thisclassifies the GrabCut as a semiautomatic segmentation tech-nique and turns the quality of the initialization and hencethe segmentation performance sensitive to the user selectionIn other words poor GrabCut initialization may lead to badfinal segmentation accuracy which might require extra userinteractions with the segmentation results for fine tuning [5]

In this work a modified GrabCut is proposed as an auto-matic segmentation technique which can segment the imageinto its natural objects without any need for the initialuser intervention Automation of GrabCut is applied usingOrchard and Bouman clustering [7] as an unsupervised clus-tering technique The selection of the Orchard and Boumanclustering is based on the empirical comparison results car-ried out in the work of [8] The paper exploits the use ofsome evaluation criteria to evaluate the discriminating powerof the automatic GrabCut with the different color spacesThe remainder of the paper is organized as follows Section 2provides a basic background on segmentation based-colorspacemodels image segmentation using GrabCut and unsu-pervised clustering techniques Section 3 explains the dif-ferent color space models Section 4 illustrates the Orchardand Bouman clustering The original GrabCut techniqueand details of its modification are explained in Section 5Experimental results are presented in Section 6 while theconclusion and future work are presented in Section 7

2 Related Work

As no common opinion has emerged about which is thebest choice for color space based image segmentation someresearch work tried to identify the best color space for aspecific task Several works [9 10] show that different colorspaces are useful for the problem of color image segmenta-tion Jurio et al [11] have carried out a comparative studybetweendifferent color spaces in cluster based image segmen-tation using two similar clustering algorithms Their studyinvolved the test of four color spaces 119877119866119861 119867119878119881 119862119872119884and 119884119880119881 in order to identify the best color representationThey obtained their best results in most cases using 119862119872119884

color space while 119867119878119881 also provided good results Businet al [4] proposed a method to automatically select a specificcolor space among classical color spaces This selection wasdone according to an evaluation criterion based on a spectralcolor analysis This criterion evaluates the quality of thesegmentation in each space and selects the best one whichpreserves its own specific properties A study of the ten mostcommon color spaces for skin color detection was presentedin [12] They concluded that 119867119878119881 is the best color spaceto detect skin in an image Another study that was appliedfor the classification of pizza toppings [13] proved that thepolynomial SVM classifier combined with 119867119878119881 color spaceis the best approach among five different color spaces Basedon a comparative study between the 119877119866119861 and 119867119878119881 models

Ruiz-Ruiz et al [14] declared that the best accuracy wasachieved with 119867119878119881 representation in order to achieve realtime processing in real farm fields for crop segmentation

GrabCut is considered one of the powerful techniquesused for color image segmentation It has been appliedto different segmentation problems such as human bodysegmentation [15ndash17] video segmentation [18] semanticsegmentation [19] and volume segmentation [20] In [17]an automatic extraction of the human body from colorimages was developed byHuThe iteratedGrabCut techniquewas used to dynamically update a trimap contour whichwas initialized from the results of a scanning detector usedfor detecting faces from images The research has somedrawbacks as the process goes through many steps anditerations in addition to being constrained to human poseswith frontal side faces A fully automatic Spatio-TemporalGrabCut human segmentation methodology was proposedby Hernandez et al [16] They developed methodologythat takes the benefits of the combination of tracking andsegmentation Instead of the initial user intervention toinitialize theGrabCut algorithm a set of seeds defined by facedetection and a skin color model are used for initializationAnother approach to segment humans from cluttered imageswas proposed by Gulshan et al in [15] They utilized thelocal colormodel basedGrabCut for automatic segmentationThis GrabCut local color model was used to refine the crudehuman segmentations they obtained In video segmentationCorrigan et al [18] extended GrabCut for more robust videoobject segmentation They extended the Gaussian mixturemodel (GMM) of the GrabCut algorithm so that the colorspace was complemented with the derivative in time of thepixelrsquos intensities in order to include temporal informationin the segmentation optimization process Goring et al [19]integrated GrabCut into a semantic segmentation frameworkby labeling objects in a given image Most recently Ramırezet al [20] proposed a fully parallelized scheme usingGrabCutfor 3D segmentation that has been adopted to run on GPUThe scheme aims at producing efficient segmentation resultsfor the case of volume meshes in addition to reducing thecomputational time

Clustering [21] the unsupervised classification of pat-terns into groups is one of the most important tasks inexploratory data analysis [22] It has a long and rich historyin a variety of scientific disciplines including anthropologybiology medicine psychology statistics mathematics engi-neering and computer science Clustering in image segmen-tations [2 23 24] is defined as the process of identifyinggroups of similar image primitives Unsupervised clusteringtechniques [25] are content based clustering where contentrefers to shapes textures or any other information that canbe inherited from the image itself

In the cases of bilabel segmentation good separationbetween foreground and background is requiredThis can beimplemented through finding clusters with a low variancesince this makes the cluster easier to separate from the othersThe selection of the Orchard and Bouman clustering tech-nique [7] is guided by Ruzon and Tomasi [26] and Chaunget al [27] in order to get tight and well separated clustersThey have worked on solving the problem of image matting







1015840))


1015840))


1015840))

1198701015840= min (11986210158401198721015840 1198841015840)

(1)


[

[

119883

119884

119885

]

]

= [

[

0412453 0357580 0180423

0212671 0715160 0072169

0019334 0119193 0950227

]

]

[

[

119877

119866

119861

]

]

(2)


[

[

119884

119880

119881

]

]

= [

[

02989 05866 01145



]

]

[

[

119877

119866

119861

]

]

(3)



119867 =

0 if Max = Min

(60∘times

119866 minus 119861

MaxminusMin+ 360

∘)


60∘times

119861 minus 119877

MaxminusMin+ 120

∘ if Max = 119866

60∘times

119877 minus 119866

MaxminusMin+ 240

∘ if Max = 119861

119878 =

0 if max = 0

MaxminusMinMax

otherwise

119881 = Max

(4)






1 and Σ

1 the covariance

matrix of 1198621





119894= 119909 isin 119862

119899 119890119879

119899119911119899le

119890119879

119899120583119899 and 119862lowast

119899= 119862119899minus 119862119894


119894





(a) (b)




1 119911

119899 119911

119873)




119873) 120572119894isin





119894

120572119873) 120572119894


119894= 1 for


119894= 0 for 119894 isin TB)














1 2 3 4 5 6 7

8 9 10 11 12 13 14

15 16 17 18 19 20 21

22 23




119894isin 1199111 119911

119899 119911

119873 119894 isin [1 119873] This means













(a) (b) (c)


(a)

(b)



1cap

11991021199101cup1199102 where119910
















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





1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in









1015840))


1015840))


1015840))

1198701015840= min (11986210158401198721015840 1198841015840)

(1)


[

[

119883

119884

119885

]

]

= [

[

0412453 0357580 0180423

0212671 0715160 0072169

0019334 0119193 0950227

]

]

[

[

119877

119866

119861

]

]

(2)


[

[

119884

119880

119881

]

]

= [

[

02989 05866 01145



]

]

[

[

119877

119866

119861

]

]

(3)



119867 =

0 if Max = Min

(60∘times

119866 minus 119861

MaxminusMin+ 360

∘)


60∘times

119861 minus 119877

MaxminusMin+ 120

∘ if Max = 119866

60∘times

119877 minus 119866

MaxminusMin+ 240

∘ if Max = 119861

119878 =

0 if max = 0

MaxminusMinMax

otherwise

119881 = Max

(4)






1 and Σ

1 the covariance

matrix of 1198621





119894= 119909 isin 119862

119899 119890119879

119899119911119899le

119890119879

119899120583119899 and 119862lowast

119899= 119862119899minus 119862119894


119894





(a) (b)




1 119911

119899 119911

119873)




119873) 120572119894isin





119894

120572119873) 120572119894


119894= 1 for


119894= 0 for 119894 isin TB)














1 2 3 4 5 6 7

8 9 10 11 12 13 14

15 16 17 18 19 20 21

22 23




119894isin 1199111 119911

119899 119911

119873 119894 isin [1 119873] This means













(a) (b) (c)


(a)

(b)



1cap

11991021199101cup1199102 where119910
















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





1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b)




1 119911

119899 119911

119873)




119873) 120572119894isin





119894

120572119873) 120572119894


119894= 1 for


119894= 0 for 119894 isin TB)














1 2 3 4 5 6 7

8 9 10 11 12 13 14

15 16 17 18 19 20 21

22 23




119894isin 1199111 119911

119899 119911

119873 119894 isin [1 119873] This means













(a) (b) (c)


(a)

(b)



1cap

11991021199101cup1199102 where119910
















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





1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




1 2 3 4 5 6 7

8 9 10 11 12 13 14

15 16 17 18 19 20 21

22 23




119894isin 1199111 119911

119899 119911

119873 119894 isin [1 119873] This means













(a) (b) (c)


(a)

(b)



1cap

11991021199101cup1199102 where119910
















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





1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a) (b) (c)


(a)

(b)



1cap

11991021199101cup1199102 where119910
















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





1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in














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





1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






1 1870 2009 545 3631 521 5857 9299 9847 9805 99242 574 279 292 291 1919 7569 9885 9863 9795 98863 731 551 390 376 537 8548 9538 9921 9931 98224 309 307 1895 308 1267 9702 9919 8813 9911 1005 375 376 420 4225 7438 8518 8933 8570 9919 96566 086 086 089 3082 2990 9717 9951 9956 7405 98207 240 233 228 120 3611 6901 9985 9976 9321 98438 108 694 108 268 268 9081 4491 9727 8869 88749 217 216 217 3143 2880 9731 9922 9924 8293 859310 281 419 423 481 569 9635 9990 9992 9739 10011 205 207 218 1499 1593 9704 9991 9959 6379 611612 493 522 497 444 829 8932 9248 9330 9681 814213 230 239 242 4231 2758 9564 9873 9903 9991 970914 256 281 306 415 428 9641 9837 9797 9479 945515 350 368 368 527 1271 9198 9928 9928 9922 884016 302 294 298 878 4966 9380 9889 9894 9956 997917 210 220 212 3589 609 9511 9907 9900 9965 811718 386 671 676 3585 8023 9106 9895 9871 9636 10019 292 3729 4578 506 640 9330 9988 6412 9029 998420 144 143 147 848 2759 9643 9895 9910 9804 988221 127 127 104 155 2274 9464 9970 9899 9818 999222 316 339 339 521 529 9373 9465 9436 8997 898523 258 321 356 325 328 9374 9513 9491 9495 9558Avg 364 549 563 1454 2131 9021 9535 9579 9354 9355SD 361 792 947 1524 2164 987 1137 784 901 929

Rate

()


120

100

80

60

40

20

0

364

549

563

1454

2131

9021

9535

9579

9354

9355

RGBYUVXYZ

CMYHSV






References










































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



color image segmentation based on different color space models using automatic grabcut

Documents