A novel image thresholding algorithm based on neutrosophicsimilarity scoreYanhui Guoa,, Abdulkadir Sengrb, Jun YecaSchool of Science, Technology & Engineering Management, St. Thomas University, Miami Gardens, FL 33054, United StatesbTechnology Faculty, Department of Electric and Electronics Engineering, Firat University, Elazig, TurkeycDepartment of Electrical and Information Engineering, Shaoxing University, 508 Huancheng West Road, Shaoxing, Zhejiang Province 312000, PR Chinaarti cle i nfoArticle history:Received 28 March 2014Received in revised form 24 June 2014Accepted 20 August 2014Available online 9 September 2014Keywords:Image thresholdingImage segmentationFuzzy setNeutrosophic setSimilarity scoreabstractImage thresholding is an important eld in image processing. It has been employed to seg-ment the images and extract objects. A variety of algorithms have been proposed in thiseld. However, these methods perform well on the images without noise, and their resultson the noisy images are not good. Neutrosophic set (NS) is a new general formal frameworkto study the neutralities origin, nature, and scope. It has an inherent ability to handle theindeterminantinformation. Noiseisonekindof indeterminantinformationonimages.Therefore, NShasbeensuccessfullyappliedintoimageprocessingandcomputervisionresearchelds. Thispaperproposedanovelalgorithmbasedonneutrosophicsimilarityscore to perform thresholding on image. We utilize the neutrosophic set in image process-ing eld and dene a new concept for image thresholding. At rst, an image is representedin the neutrosophic set domain via three membership subsets T, I and F. Then, a neutro-sophic similarity score (NSS) is dened and employed to measure the degree to the idealobject. Finally, an optimized value is selected on the NSS to complete the image threshold-ingtask. Experimentshavebeenconductedonavarietyof articial andreal images.Several measurements areusedtoevaluatetheproposedmethods performance. Theexperimental results demonstrate that the proposed method selects the threshold valueseffectively and properly. It can process both images without noise and noisy images havingdifferentlevelsof noiseswell. It willbe helpfultoapplicationsinimageprocessingandcomputer vision. 2014 Elsevier Ltd. All rights reserved.1. IntroductionImage thresholding, one of the simple image segmenta-tion procedures, is a crucial step for several image-process-ing applications such as object detection, shape recognition,andoptical characterrecognition[1]. Intheimage thres-holding process, a threshold value is selected, and the pixelsontheimagesareclassiedintobackgroundor objectsaccording to their values. Image thresholding can convertthe gray level images into binary ones [2]. Thresholding isquite efcient whenthe object pixels andbackgroundpixelshave distinct gray level distributions. Furthermore, it is easyto be implemented and usually be run fast [3,4].A variety of algorithms have been proposed. Generally,image thresholding methods are classied into two groupsbased on the criteria to select the threshold value: globaland local methods [5]. Global methods select the thresholdvalues according to the characteristics of the entire images,and local ones adopt threshold values using the local infor-mationontheimages. Thresholdvalue selectionmethodhttp://dx.doi.org/10.1016/j.measurement.2014.08.0390263-2241/ 2014 Elsevier Ltd. All rights reserved.Corresponding author.E-mail addresses: yguo@stu.edu (Y. Guo), ksengur@rat.edu.tr(A. Sengr), yehjun@aliyun.com (J. Ye).Measurement 58 (2014) 175186ContentslistsavailableatScienceDirectMeasurementj our nal homepage: www. el sevi er. com/ l ocat e/ measur ementbased on image histogram is a kind of the global methods [6].Forahighcontrastimage, thehistogramhastwodistin-guished peaks, and a wide valley between the two peaks.Thethresholdvalueisselectedthevalueinthevalley.However,thehistogram basedmethods failtondapropervalue to segment the image on a low contrast image becausethe histogram does not have distinguished peaks and valleys.Avarietyof methodshavebeenpresentedtoselectthethresholds based on histogram and fuzzy logic [714].Afuzzybasedimagethresholdingschemewas pro-posed by Pal and Rosenfeld [7]. The authors used the fuzzycompactness by using the S-function for membership eval-uation. Huang and Wang proposed an efcient fuzzy thres-holding methodbased onYagers measure whichis ameasure of fuzziness depending on the relationshipbetweenthefuzzyset anditscomplement [8,9]. ChairaandRay[10] usedtheGammamembershipfunctiontocomputethemembershipvaluesof thepixels, andpro-posed the fuzzy divergence for image thresholding. Ramaret al. proposed the neural networks for selecting the opti-mum threshold value using fuzzy measure [11]. Cheng andChenusedfuzzyhomogeneityandfuzzyco-occurrence(a)Artificialimages with Gaussian noise.(b) Result by the NS method.(c) Result by the NSS method.Fig. 1. Performance comparison on a high contrast articial noisy image.(a) Low contrast artificial image with Gaussian noise.(b) Result by the NS method.(c) Result by the NSS method.Fig. 2. Performance comparison on a high contrast articial noisy image.176 Y. Guo et al. / Measurement 58 (2014) 175186matrix for image thresholding [12]. The method [12]employed the homogeneity vectors and the fuzzy member-shipfunction, andextractedthefeatureof animagetodeterminefuzzyregions. Thefuzzyentropyvalueswereutilizedtodeterminethethresholdsforsegmentingtheinput images. Tizhoosh proposed a thresholding techniquebased on ultra-fuzzy sets [13]. The ultra-fuzzy set was usedtoremovethevaguenessintheimage. Chengetal. pro-posed a two dimensional fuzzy entropy method for obtain-ingthebestthresholdvalue[14]. Theproposedmethodinvolvedfuzzypartitioningonatwo-dimensional histo-gramwherefuzzyentropywasdened. Finallyageneticalgorithmwasemployedtoobtaintheoptimalthresholdvalue.In[15], Xiaoet al. presentedathresholdingmethodusing an articial bee colony (ABC) algorithm and entropyfunction. TheABCsearchedthemaximumvalueof theentropy, and the optimal thresholds were determinedbased on the maximum. Xiong et al. proposed a thresholdselectionmechanismfor theradar images thresholdingwhich combines the characteristics of two different mea-sures using the Markov random eld model [16]. In [17],He et al. presented a threshold method using a two-dimen-sional histogram and multi-resolution analysis. Themethod determined the optimal threshold value usingthe spatial correlationof graylevel andthe exibility,and searched the threshold value via multi-resolutionway. Junetal. proposedatwo-dimensional Tsallissym-metric cross entropy for image thresholding [18]. Thetwo-dimensional Tsallis symmetric cross entropy wasdened, and a recursive algorithm was used to search theoptimal thresholdvector. Afuzzyentropymeasureonatwo-dimensional histogrammethod was proposed in[19]. The image was separated into several different gridswithdifferentdensities. Thentheintensityintheimageand the average intensity of the local neighborhood wereused to build a two-dimensional histogram. A multi-thresh-old method was presented based on the maximum fuzzyentropy principle and the two-dimensional histogram. Theparameters of the entropy function were tuned via agenetic algorithm. In [20], Bustince et al. dened an igno-rancefunctionandusedit toobtainathresholdvalue.Measurementswereconstructedusingt-normsandautomorphism. The degree of ignorance was employed todescribethebackgroundandobjects. Basedontheigno-rance degree, the threshold is obtained from the interval-valued fuzzy set having the lowest associated ignorance.However, theabovemethodssufferfromndingtheoptimal threshold value when the input images have noise.Especially, under the low SNR levels the above-mentionedmethods achievementsdropconsiderably. Toovercomethe limitations of the above methods, we proposed a neu-trosophicset basedimagethresholdingschemeforef-cient bi-level segmentation. More specically, theproposedmethodusestheneutrosophicsimilaritymea-sures for determining the optimal threshold value.Neutrosophy is a new kind of generalizations of dialec-tics, andit studies the neutralities origin, nature, andscope[21]. It represents everyentityhAi, theoppositehAnti-Ai, andtheneutralities hNeut-Ai thatisneither hAinor hAnti-Ai.The traditional fuzzy set utilizes a membership to rep-resent the degree belonging to a set. When the fuzzy mem-bership value is uncertain and vague, it is challenging to bedenedusingacrispvalue[22]. Insomesituations, wehave to consider both the membership and the indetermi-nacy of the membership.In the neutrosophic set (NS), each entity is depicted viathree memberships: truth, indeterminacy and falsity mem-berships. Thischaracteristicisessential tosuchapplica-tions as information fusion where data might have adegree of uncertainty.Theimagethresholdingmethodsbasedonthetradi-tional fuzzy set can be affected by noise severely. This papernewly develops a neutrosphic set approach for image thres-holding. First, animageismappedintotheNSdomain.Then, a novel similarity measurement, neutrosophic simi-larityscore, isdenedtomeasurethepixels belongingdegree to the object on the image. Finally, the image is per-formed thresholding using the neutrosophic similarityscore. The experiments on synthetic images having differ-entlevelsofnoiseandnoisyrealworldimagesarecon-ducted to evaluate the proposed approachs performance.Thepaperisorganizedasfollows. Section2describesthe proposed method, Section 3 discusses the experimentalresults and comparisons, and the conclusions are drawn inSection 4.Fig. 3. The relation between SNR and ME. *: NSS method, O: NS method.Fig. 4. The relation between SNR and FOM. *: NSS method, O: NS method.Y. Guo et al. / Measurement 58 (2014) 175186 1772. Proposed method2.1. Neutrosophic similarity scoreA neutrosophic set can be dened under different crite-ria as: let A fA1; A2; . . . . . . ; Amg be a set of alternatives inneutrosophic set, and C fC1; C2; . . . . . . Cng be a set of cri-teria. The alternative Aiat Cjcriterion is denoted asTCjAi; ICjAi; FCjAin o=Ai, where TCjAi, ICjAi and FCjAiare the membershipvalues to the true, indeterminacyand false set at the Cj criterion.Asimilaritymeasurementisproposedtoevaluatethesimilaritydegreebetweentwoelementsinneutrosophicset under multicriteria [23]:SCjAm; AnTCjAmTCjAn ICjAmICjAn FCjAmFCjAnT2CjAm I2CjAm F2CjAmq T2CjAn I2CjAn F2CjAnq 1In multi-criteria environment, the concept of ideal elementcanbe usedtoidentifythe best alternative. The idealalternative A*is denoted as: fTCjAi; ICjAi; FCjAig=Ai.The similarity to the ideal alternative is computed as:SCjAi; A TCjAiTCjA ICjAiICjA FCjAiFCjAT2CjAi I2CjAi F2CjAiq T2CjA I2CjA F2CjAq2An image is dened in the NS as: let U be a universe, BPbe a bright pixel set in U, and an image Im described usingNS is called neutrosophic image INS. The neutrosophicimageINSisdepictedusingsubsetsT, I andF. ApixelinINSisdenotedasPNS(T, I, F), anditbelongstothebrightpixel set BP in the means as: it is T true in the bright pixelset, I indeterminate, and F false. The range of the values inT, I is in [0 1].Accordingtothedenitionof neutrosophicimage, apixel Px; y is interpreted in the neutrosophic set domain:PNSx; y fTx; y; Ix; y; Fx; yg. Tx; y, Ix; yandFx; yrepresent memberships belonging to bright pixel set, inde-terminate set and non-bright pixel set, respectively. At theintensity criterion, they are dened as: (a)Lenaimagewith differentGaussiannoise level: variance: 0, 10, 20 (b)Thresholdingresultsof the NS method. (c)Thresholdingresultsof the NSS method. Fig. 5. Comparison results on Lena image.178 Y. Guo et al. / Measurement 58 (2014) 175186TCgx; y gx; y gmingmax gmin3ICgx; y 1 Gdi; j GdminGdmax Gdmin4FCgx; y 1 TCgx; y 5where g(x,y) and Gd(x,y) are the intensity value and gradi-ent value at the position of (x,y) on the image.Then, asimilarityscoreis calculatedtoidentifythedegree to the ideal object under intensity condition.SCgPx; y; ATCgx; yTCgA ICgx; yICgA FCgx; yFCgAT2Cgx; y I2Cgx; y F2Cgx; yq T2CgA I2CgA F2CgAq6The similarity value is sensitive to noise on image. If it isused for image thresholding, the noisy regions on imageswill be labeled into awrong group. In order to make thethresholding results robust to noise, we propose two newcriteria, local mean intensity criterion Cm and local homo-geneitycriterionCh, andthencalculatetheneutrosophicsimilarity score under them.Theneutrosophicset under thelocal meanintensitycriterion Cm is dened as:TCmx; y gmx; y gmmingmmax gmmin7gmx; y 1w wXxw=2mxw=2Xyw=2nyw=2gm; n 8(a) Woman image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 6. Comparison results on Woman image.Y. Guo et al. / Measurement 58 (2014) 175186 179ICmx; y 1 Gdmi; j GdmminGdmmax Gdmmin9FCmx; y 1 TCmx; y 10where gm(x,y) and Gd(x,y) are the intensity value and gradi-ent magnitude value at the position of (x,y) on the imageafter mean lter processing. gm min and gm max are minimumand maximum intensity values, respectively. Gdm min and Gdmmax are minimum and maximum of the gradient value in theimage. Meanlteringisknownasthesimpleandintuitiveprocedure that has been used in image processing for smooth-ingtheimagesandnoisereduction. Theprocedureofmeanltering is replacing each pixel0s intensity value in an imagewith the mean value of its neighbors and itself. The mean ltereliminatesnoisesbyreplacingit usingitssurroundings. Inaddition, homogeneity is largely related to the local informa-tionextractedfromanimageandreects howuniformaregion is. It plays and important role in image segmentation.The neutrosophic set under the local homogeneity crite-rion Ch is also dened as:TChx; y Hx; y HminHmax Hmin11IChx; y 1 Gdhi; j GdhminGdh max Gdh min12FChx; y 1 TChx; y 13Hx; y TEMgx; y 14whereH(x,y)isthehomogeneityvalueat(x,y), whichisdepictedas the lteringresult withthe texture energymeasures(TEM)lters[31]. Gdh(x,y)isthegradientvalueon H(x,y). Hminand Hmaxare minimumand maximumhomogeneity values respectively. In addition, Gdh minandGdh maxareminimumandmaximumof thegradientvalue.Usingthesameway, anothertwoNSSvaluestotheideal alternative are calculated as:The value of A*under three criteria are same as:fTCjAi; ICjAi; FCjAig=Ai f1; 0; 0g=A .(a) Panda image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 7. Comparison results on Panda image.SCaPx; y; A TCax; yTCaA ICax; yICaA FCax; yFCaAT2Cax; y I2Cax; y F2Cax; yq T2CaA I2CaA F2CaAq 15SChPx; y; A TChx; yTChA IChx; yIChA FChx; yFChAT2Chx; y I2Chx; y F2Chx; yq T2ChA I2ChA F2ChAq 16180 Y. Guo et al. / Measurement 58 (2014) 175186Intheproposedmethod, weusetheaveragevalueofSCg, SCmandSChasthenalneutrosophicsimilarityscorefor the future application, such as image thresholding:NSx; y SCgx; y SCax; y SChx; y3172.2. Image thresholding based on NSSAftertheneutrosophicsimilarityscoreisdetermined,Otsus method is employed to obtain the optimized valuefor thresholding [24]. It selects anoptimumthresholdvalue by minimizing the within-class variance of the objectand background.The variance is used to dene a function as:rt c1t r1t c2t r2t 18where r(t) is the sum of variances of the background andobjectpixelsseparatedbythethresholdt. r1(t) andc1(t)are the variance and probability of pixels whose intensitiesare less than the threshold t, while r2(t) and c2(t) are thevariance and probability of pixels whose values are greaterthan t.Theoptimizedthresholdt*withtheminimizationofr(t) can separate the pixels into the foreground and back-ground groups.t arg minmint6t6maxtrt 19The data whose value is greater than t*will be groupedinto object set Obj, and others are background set Bkg.Obj ftjt > tg 20Bkg ftjt 6 tg 21(a) Cameraman image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 8. Comparison results on Cameraman image.Y. Guo et al. / Measurement 58 (2014) 175186 181In the proposed thresholding method, we calculate theprobabilitiesandvariancesof thetwoclassesusingtheNSS value on the image and select the tNSS*value with min-imized sum of variances of background and objects as theoptimized threshold value. The image will be segment intobackground and object sets using the optimized thresholdvalue on NSS.rNSSt cNSS1t rNSS1t cNSS2t rNSS2t 22tNSS arg minmintNSS6t6maxtNSSrtNSS 23The steps in the algorithm can be summarized as follows:Step 1: Convert the image into NS domain.Step 2: Compute the NSS values under three conditions.Step 3: Calculate the probabilities and variances of thetwo classes using the NSS values.Step 4: Select the value with minimized sum of the twoweighted variances on NSS values as the thresh-old value.Step 5: Transfer the image into binary image using thethreshold value on NSS.3. Experimental results and discussionsWe have tested the proposed algorithm using differentimages, and compared its performance with those of newlydeveloped algorithms. In the experiments, we compare theNSS method with a newly published thresholding methodbasedonneutrosophic set (NS) [25] whichperformedbetterthresholding results than those of Otsu method [24], Parzenwindowtechnique[26], theminimal error thresholding(MET)algorithm[27]andaentropybasedapproach[28].3.1. Performance on articial imagesWeusearticial imagestocomparetheNSSmethodandNSmethodvisually, andthenevaluatetheirresults(a) Discover image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 9. Comparison results on Discover image.182 Y. Guo et al. / Measurement 58 (2014) 175186quantitatively. In [25], several articial images having twointensities (one is 0 and another is 255) and impaired bydifferent levels of noise were employed to evaluation theperformance of the NS method. In Fig. 1, the rst columnhas two same articial images as in [25] having Gaussiannoise, whose mean values are 0 and standard variance val-ues 25.5 and 178.5, respectively. The second and third col-umnsdisplaythethresholdedresultsbytheNSmethodand NSS method. Fromthe results in Fig. 1, the NSS methodachieved the same performance as the NS algorithm.Toidentifytheperformancedifferenceof theNSandNSS methods, we employ a new articial image with lowercontrast anddifferent levelsof Gaussiannoise. Anothersyntheticimagehavingtwograylevels(64and128) isimpaired bydifferentlevelsofGaussiannoise totesttheperformanceof twothresholdingmethods. Fig. 2(a) aresynthetic images having Gaussian noise, whose meanvalues are 0 and standard variance values are 15 and 30,respectively. Figs. 2(b)and(c)aretheresultsbytheNSand NSS methods, respectively.TheresultsonthesecondsyntheticimagewithlowcontrastshowtheNSSmethodperformsbetterthanNSmethod on the synthetic images with two different noiselevels. AlotofpixelsinFig. 2(b)areidentiedinwrongclasses, while they are classied correctly by NSS methodin Fig. 2(c).To evaluate the thresholding results for articial imagesquantitatively,we utilize a metric: misclassication errormeasure (ME) [26,29], to measure the thresholding perfor-mances. The MEdepicts the percentage of backgroundpoints wrongly grouped to foreground set, and objectpoints wrongly grouped to background set [26,29]:ME 1 jB0 \ BTj jF0 \ FTjjB0j jF0j24(a) Blood image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 10. Comparison results on Blood image.Y. Guo et al. / Measurement 58 (2014) 175186 183where F0 and B0 are the object and background sets on thegroundtruthimage. FTandBTaretheobject andback-ground sets on the result image. || denotes the elementsnumber in the set. The ME denes a metric of themisgroupedpointsbetweenthegroundtruthimageandthe test image.Another metric gure of meritFOM proposed by Pratt[30] is alsoutilizedtoevaluatethedifferencebetweenthemethods results withtheideal thresholdingresultquantitatively:FOM 1maxNI; NAXNAk111 bd2k25where NI and NA are the numbers of the object points andtheideal object pixels, respectively. d(k) isthedistancefrom the kth actual point to the nearest thresholding point.bisaconstant, andsetas1/9inPrattspaper[30]. Thegreater the FOM, the better the thresholding results are.The quality of noisy images can be measured using theSNR (signal to noise ratio):SNR 10logPH1r0PW1c0im2r; cPH1r0PW1c0 imr; c imnr; c2" #26where imn(r, c) and im(r, c) are the intensities of pixel (r, c)in the noisy and original images, respectively.The values of ME are plotted at different SNR levels inFig. 3. It demonstrates theNSSmethodarchives lowerMEs at all SNRs. All ME values of the NSS method are smal-ler than 0.015, and all values of NS method are higher thanthose of the NSS method. The NSS algorithm achieves the(a) Rice image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 11. Comparison results on Rice image.184 Y. Guo et al. / Measurement 58 (2014) 175186optimumperformancewithME = 0.0065whichisnearlyequal tozerowhenSNRis20.1144 dB, whileNSobtainsthe optimum value with ME = 0.0068. Meanwhile, the val-ues of FOM of NSS are bigger than those of NS at most SNRlevels. The average value of FOM of NSS and NS are 0.953and 0.8517, respectively (see Fig. 4).3.2. Performance on real world imagesA great number of real world images are employed tomeasure the NSS methods performance. Eight representa-tive images are selected to demonstrate the effectivenessandrobustness of theNSSmethodagainst thevariousnoise levels. These images are Lena, Woman, Panda,Cameraman, Discover, Blood, Rice and Coins,respectively. InFigs. 512, therstrowliststheoriginalimages without noise and the results of NS and NSSmethod on them. The second and third rows demonstratetheresultsonthenoisyimageswithdifferent Gaussiannoise variances. The experimental results demonstrate thatthe results by NSS method exhibits visually better qualitythanthoseofNSmethod. TheresultsbyNSmethodareaffectedseverelybythenoiseonthereal images. Theresults of the NSS method eliminate noise effect and mostpixels are segmented into the right groups.The thresholding results of the Lena image are givenin the second and the third columns of the Fig. 5. It is obvi-ousthattheresultsobtainedundertherstandsecondnoise variance are better than the NS method. The shoul-der, faceandthehatregionsarerecoveredcorrectly. Onthe other hand, the NS based method yields several wrongsegmentations especially in the shoulder and the faceregions. It is worth mentioning that the both NSS and NShave several wrong segmentations for the third noise level.Especially both methods have misclassied pixels thebackground region and the shoulder and face of the Lena.However, byvisualinspection, theproposedmethodhasbettersegmentationthantheNSmethodwhenwecon-sider theoverall results. Similar results canbeseeninthe Woman image, which is depicted in Fig. 6. The face,hair, background regions are thresholded correctly by theNSS method for the rst and second noise levels. Inaddition, the proposed NSS method can obtain reasonablesegmentations even for the third noise level. For the sec-ondandthirdnoiselevelstheNSbasedmethodyieldsmany misclassications inthe face andbackgroundof(a) Coins image with different Gaussian noise level: variance: 0, 10, 20(b) Thresholding results of the NS method.(c) Thresholding results of the NSS method.Fig. 12. Comparison results on Coins image.Y. Guo et al. / Measurement 58 (2014) 175186 185the Womanimage. In Fig. 7, there are two pandas in agrassbackground. Intherst noisevariancelevel, bothNS and NSS methods obtain similar segmentation results.The background and the pandas are correctly segmented.For the second and third noise variance levels, the imagebecomes more complicatedand the NS methodyieldsmanymisclassiedpixelsinthebackground. Therearealso several numbers of misclassied pixels on the pandaregions. The superiority of the proposed NSS method canbeseenintherestof theexperiments. IntheCamera-man, Discover, Blood, Rice andCoins images, itis obvious that our proposal obtains almost all the groundtruth segmentations. Especially for the Discover, Rice,Blood andCoins imagestheobtainedresultsfor allnoise variance levels are better than the NS algorithm. Onlythere are several misclassied pixel in the grass region ofthe Cameraman image.Fromtheexperimentsonthearticialandrealworldimages, wecanmakeaconclusionthattheNSSmethodisfeasibletoselectbetterthresholdvaluesforboththeclear images without noise and the noisy images with dif-ferent levels.4. ConclusionsThis paper presents anewimagethresholdingalgo-rithmusingneutrosophicsimilarityscore. Theimageisdepicted in neutrosophic set via three subsets. Then, a neu-trosophic similarity score is dened to measure the degreetotheobject pixelsontheimage. Finally, anoptimizedvalue is selected on the NSS to perform image threshold-ing. TheexperimentalresultsshowthattheNSSmethodcan obtain the thresholds properly and effectively. It is abletoprocessbothimageswithoutnoiseandnoisyimageswithdifferentlevelsofnoises. Thisadvantagewill bringthe proposed method into more applications in theresearchareas suchas imageprocessingandcomputervision.References[1] J.S. Weszka, A survey of threshold selection techniques, Comp.Graph. Image Process 7 (2) (1978) 259265.[2] P.K. Sahoo, S. Soltani, A. Wong, A survey of thresholding techniques,Comput. Vision Graph. Image Process 41 (2) (1988) 233260.[3] M. Sezgin, B. Sankur, Survey over image thresholding techniques andquantitative performance evaluation, J. Electron Imag. 13 (1) (2004)146168.[4] N.R. Pal, S.K. Pal, Areviewon image segmentation techniques,Pattern Recogn. 26 (9) (1993) 12771294.[5] R.C. Gonzalez, R.E. Woods, Digital Image Processing, seconded.,Prentice Hall, 2002.[6] H.D. Cheng, X.H. Jiang, Y. Sun, J.J. Wang, Color image segmentation:advance and prospects, Pattern Recogn. 34 (2001) 22592281.[7] S.K. Pal, A. Rosenfeld, Image enhancement and thresholding byoptimization of fuzzy compactness, Pattern Recogn. Lett. (7) (1988)7786.[8] L.K. Huang, M.J. Wang, Image thresholding by minimizing themeasure of fuzziness, Pattern Recogn. 28 (1) (1995) 4151.[9] R.R. Yager, Onthemeasureof fuzziness andnegation, Part 1:membership in the unit interval, Int. J. Gen Syst 5 (1979) 221229.[10] T. Chaira, A.K. Ray, Segmentationusingfuzzydivergence, PatternRecogn. Lett. 12 (24) (2003) 18371844.[11] K. Ramar, S. Arumugam, S.N. Sivanandam, L. Ganesan, D.Manimegalai, Quantitativefuzzymeasuresforthresholdselection,Pattern Recogn. Lett. 21 (1) (2000) 17.[12] H.D. Cheng, C.H. Chen, H.H. Chiu, Imagesegmentationusingfuzzyhomogeneity criterion, Inform. Sci. 98 (1997) 237262.[13] H.R. Tizhoosh, Image thresholding usingtypeIIfuzzysets, PatternRecogn. 38 (2005) 23632372.[14] H. D Cheng, Y.H. Chen, X.H. Jiang, Thresholding using twodimensional histogramandfuzzy entropy principle, IEEE Trans.Image Process. 9 (4) (2000) 732735.[15] Y. Xiao, Y. Cao, W. Yu, J. Tian, Multilevel threshold selection basedon articial bee colony algorithm and maximum entropy for imagesegmentation, Int. J. Comput. Appl. Technol. 43 (2012) 343350.[16] B. Xiong, Q. Chen, Y. Jiang, G. Kuang, A threshold selection methodusingtwoSARchangedetectionmeasuresbasedonthemarkovrandomeldmodel, IEEEGeosci. RemoteSens. Lett. 9(2)(2012)287291.[17] S. He, J. Ni, L. Wu, H. Wei, S. Zhao, Imagethresholdsegmentationmethodwith2-Dhistogrambasedonmulti-resolutionanalysis,Comput. Sci. Ed. ICCSE 09 (2009) 753757.[18] Y. Jun, Z. Jun, Y. Xuefeng, Two-dimensional Tsallis symmetric crossentropy image thresholdsegmentation, informationscience andengineering (ISISE), in: 2012International Symposiumon, 2012pp.362366.[19] J. Lan, Y. Zeng, Multi-threshold image segmentation using maximumfuzzyentropybasedonanew2Dhistogram, Optik-Int. J. LightElectron Opt. 124 (18) (2013) 37563760.[20] H. Bustince, M. Pagola, E. Barrenechea, J. Fernndez, P. Melo-Pinto, P.Couto, J. Montero, Ignorance functions. An application to thecalculationof thethresholdinprostateultrasoundimages, FuzzySets Syst. 161 (1) (2010) 2036.[21] F. Smarandache, A Unifying Field in Logics Neutrosophic LogicNeutrosophy Neutrosophic Set,Neutrosophic Probability, third ed.,American Research Press, 2003.[22] H. Wang, R. Sunderraman, F. Smarandache, Y.Q. Zhang, IntervalNeutrosophic Sets and Logic: Theory and Applications in Computing:Innite Study, 2005.[23] J. Ye, Multicriteriadecision-makingmethodusingthecorrelationcoefcient under single-valuedneutrosophic environment, Int. J.Gen Syst 42 (4) (2013) 386394.[24] N. Otsu, A threshold selectionmethodfromgray-levelhistograms,IEEE Trans. Syst. Man Cybernet. 9 (1) (1979) 6266.[25] H.D. Cheng, Y. Guo, A newneutrosophic approach to imagethresholding, New Math. Nat. Comput. 4 (3) (2008) 291307.[26] S.T. Wang, F.L. Chung, F.S. Xiong, A novel image thresholding methodbasedonParzenwindowestimate, PatternRecogn. 41(1)(2008)117129.[27] J. Kittler, J. Illingworth, Minimumerror thresholding, Pattern Recogn.19 (1) (1986) 4147.[28] J.N. Kapur, P.K. Sahoo, A.K.C. Wong, Anewmethodforgray-levelpicturethresholdingusingtheentropyofthehistogram, Comput.Vision Graph. Image Process. 29 (3) (1985) 273285.[29] W.A. Yasnoff, J.K. Mui, J.W. Bacus, Error measures for scenesegmentation, Pattern Recogn. 9 (4) (1977) 217231.[30] W.K. Pratt, DigitalImageProcessing, JohnWiley&Sons, 1978. pp.429432.[31] K. Laws, Textured Image Segmentation, Ph.D. Dissertation,University of Southern California, January 1980.186 Y. Guo et al. / Measurement 58 (2014) 175186