SEGMENTATION USING NEW TEXTURE FEATURE

Transcription

1 SEGMENTATION USING NEW TETURE FEATURE S.Md.Mansoor Roomi 1 M.Mareeswari 2 G. Maragaham 3 1 Assisan Professor, Dearmen of Elecronics and Communicaion Engineering, TCE, Madurai 2 Projec Associae, Dearmen of Elecronics and Communicaion Engineering, TCE, Madurai 3 Assisan Professor, Dearmen of Elecronics and Communicaion Engineering, Universiy College of Engineering, Dindigul ABSTRACT Color, exure, shae and luminance are he rominen feaures for image segmenaion. Texure is an organized grou of saial reeiive arrangemens in an image and i is a vial aribue in many image rocessing and comuer vision alicaions. The objecive of his work is o segmen he exure sub images from he given arbirary image. The main conribuion of his work is o inroduce NEW exure feaure descrior o he image segmenaion field. The NEW exure descrior labels he neighborhood ixels of a ixel in an image as N,W,NW,NE,WW,NN and NNE(N-Norh, W-Wes).To find he redicion value, he gradien of he inensiy funcions are calculaed. Eigh comonen binary vecors are formed and comared o redicion value. Finally end u wih 256 ossible vecors. Fuzzy c-means clusering is used o segmen he similar regions in exural image Exensive exerimenaion shows ha he roosed mehodology works beer for segmening he exure images, and also segmenaion erformance are evaluaed.. KEYWORDS NEW exure feaure; Fuzzy c-means clusering; brodaz daase 1. INTRODUCTION Wih advances in digial imaging, digial hoograhy collecion here is an exlosive growh of image daabases. Searching of an image in such large collecions ake enormous amoun of ime. Tyically, he simles way o manage he large image collecions is convenional daabasemanagemen sysems like relaional daabases or objec-oriened daabases. In his ye of sysems, images are annoaed wih keywords. Bu hese kinds of sysem have failed when he collecion of images growing larger. To overcome his roblem, Conen Based Image Rerieval (CBIR) has been emerged. In CBIR, images are indexed by heir own visual conens. Image classificaion is necessary for CBIR o classify he visual conens of images. Feaure exracion is one he echniques used o classify he images. Feaures such as exure, color and shae are commonly used in sae of ar mehods. In he lieraure, mos of he aenion has been focused on he exure feaures. Texure is a key characerisic of an image hels o idenify objecs or regions of ineres in he image. Texure gives informaion abou he saial arrangemen of inensiies in an image (Jain 1998). Haralick e al (1973) menioned ha exures can be fine, coarse, smooh, riled, irregular, lineaed ec. An examle of exures is shown Figure 1. DOI: /acij

2 Figure 1 Samle exure images Source: Brodaz exure daase For exure feaure exracion various mehods like model based, saisical and single rocessing mehods are used (Zhang e al 2006). Model based mehods describe he exure images based on he robabiliy disribuion. In model based echnique, number of random field models such as fracals, auoregressive models, fracional differencing models, and Markov random fields (MRF) have been used for modeling and synhesis of exure. MRFs are widely used because i yields a local and economical exure descriion. Zi (1995) defines a MRF on a discree laice wih resec o a neighbourhood sysem by local condiional robabiliies. Efficien arameer esimaion scheme is mus for a model based aroach o be successful. In saisical mehods, exures are characerized using saisical measures, such as co- occurrence or saial auocorrelaion of he gray levels. Haralick e al (1973) roosed a Grey Level Co occurrence Marices (GLCM) o exrac geomerical feaures like energy, enroy, correlaion, ec. By esimaing air wise saisics of ixel inensiy, GLCM is buil from he image. The roblem is co occurrence marix is calculaed based on he dislacemen vecor. In saisical mehod, when he order of saisics (k) is large (k >2) i is hard o handle because enormous amoun of daa are involved. Mos recenly he signal rocessing mehods using muli resoluion and muli channel are inroduced for exure analysis and classificaion. In his mehod, by using bank of filers such as Gabor (Jain e al 1997), neural nework (Karu e al 1996), wavele based filers (chen e al 1997), a exural image is decomosed ino feaure images. Smih e al (2005) roosed a mehod o exrac he exure feaure from he mean and variance of wavele sub bands. Laer, o erform exure analysis wavele Transform ogeher wih KL exansion and kohenon mas was develoed by Cross e al (2007), Bourchani e al (2005) combine he wavele ransform wih co occurrence marix o ge he advanages of saisics based and ransform based exure analysis. Tree srucured wavele ransform and Gabor wavele ransform is inroduced by W.Y Ma e al (2008) o erform exure image annoaion. As a resul, wih hel of se of well seleced filers a high dimensional exural aern can be exraced. Therefore, he major issue of his mehod is he selecion of good se of filers. 40

3 In his aer we concerned wih he ask of roosing a mehod o comue NEW exure descrior. Conex Adaive Lossless Image Comression (CALIC) scheme is used for image comression. This scheme calculaes a redicion value of a curren ixel based on he neighbourhood ixels which forms eigh comonen binary vecor afer ha a conex is consruced for comressing he given image. The auhor menions ha he eigh comonen binary vecor is a exure descrior [ ]. We have used his exure descrior for image segmenaion wih he hel of Fuzzy C-Means algorihm. The exure is deermined by labeling he neighbourhood ixels as Norh, Eas and Wes. So, we named he exure feaure as NEW exure descrior. The main conribuions of his aer include: A novel NEW exure feaure is inroduced o exrac he exure informaion from he exure image A mehodology o segmen and classify he exure image ino four regions Fuzzy C-Means algorihm is used o segmen he exure image based on NEW exure feaure Exensive exerimens are carried ou o rove he efficiency of he roosed mehodology The res of his aer is organized as follows: Secion II exlains abou he roosed mehodology. The exerimen resuls are reored in Secion III, and Finally, secion IV resens conclusion. 2. PROPOSED METHODOLOGY 2.1. Feaure exracion using NEW exure descrior This secion discusses abou a novel framework o exrac he NEW exure descrior for classificaion of images. Figure 3 shows he overview of he roosed NEW feaure exracion. Consider a ixel marked as in an image label, he o ixel of is denoed as N, lef ixel as W, likewise all he neighborhood ixels of are labeled as shown in Figure 3 Figure 2 Labeling he neighbors of ixel Fig.2 labeling he neighbors of ixel The informaion abou he neighborhood ixels are quanified by forming he eigh comonen vecor as N, W, NW, NE, NN, WW, 2N-NN and 2W-WW. Deending on he verical and horizonal edges of he neighborhood of he ixel may give he bes redicion value denoed as. How close he redicion value is urely deends on he surrounding exure. The informaion abou he kind of boundary can be calculaed from he gradiens of he inensiy as shown in equaion 1 and equaion 2. dh= W-WW + N-NW + NE-N (1) dv= W-NW + N-NN + NE-NNE (2) 41

4 The redicion value is comued from he inensiy gradiens dh and dv. If he value of dh is much higher han he value of dv (dh >>dv), here will be a horizonal variaion. So N is seleced as he redicion value. Verical variaion will be found, if he value of he dh is much smaller han he value of dv (dh <<dv), and he redicion value is assumed o be W. If he difference of dh and dv is moderae hen he weighed average of neighboring ixels is assigned o redicion value. Pseudo code for finding redicion value is given below: if dh-dv > 80 =N else if dv-dh >80 =W else { } = (N+W)/2+(NE-NW)/4 if dh-dv >32 =( else if dv dh>32 =( else if dh dv>8 =(3 else if dv dh>32 =(3 +N)/2 +W)/2 +N)/4 +W)/4 Fig.3 Overview of NEW exure descrior exracion 42

5 Afer finding he redicion value, eigh comonen binary vecors is consruced. To do so, each comonen of eigh comonen vecor formed by neighbors of x is comared wih he redicion value. If he value of comonen is less han he hen relace he value wih 1 else 0. Thus, eigh comonen binary vecors creaed is called NEW exure descrior Texure image segmenaion In his aer, Fuzzy C-Means algorihm [18] is used o cluser he image ino mulile segmens using NEW exure feaure. Fuzzy C-Means clusering algorihm segmens he image based on he disance beween cluser cener and oher feaure domain.this feaures are clusered by minimizing he objecive funcion (3) J FCM = r n i= 1 = 1 u q i d 2 i (3) Where r reresens number of clusers, i=1, 2,,r and is he number of ixels =1,2.n, di u is he disance and i is he udaion of cluser ceners. Fig. 4. Shows he roosed aroach o segmen he image ino mulile regions. Iniially, he inu image is divided ino 32x32 overlaing blocks. NEW exure feaure is exraced from each block o segmen he region. This segmenaion algorihm requires user ineracion o assign number of clusers (K) for segmening he image. To circumven his roblem, his aer uses an auomaic assigning of number of labels for segmenaion algorihm using he number of eaks esimaed from he hisogram of he inu exure image afer smoohing he hisogram using Gaussian filer[17]. Inu image Hisogram Calculaion Divide he image ino32x32 blocks Number Of Peaks deerminaion o find K Exrac NEW exure feaure from each block FCM clusering Clusered regions Fig.4 Flow diagram for image segmenaion 43

6 Hisogram of he grayscale image is calculaed as H ( x) δ ( x, = MN i= 1 x i ) and x [ x0, xl 1] (4) 1 = 0 δ ( x, y) Where H s ( ) = K( )* H ( ) (5) (6) Where K is Gaussian smoohing kernel [19] and H is he hisogram of he image and =0, 1, 2, Exerimenal resuls and discussion In his secion, he efficiency of he roosed mehodology is analyzed and resuls are shown Daa descriion In his area 12 differen exure images are aken by brodaz daase. The images are available in online. Fig.5 shows he combinaion of 4 differen exure images. Fig. 5. Texure image from brodaz daase 2.5. Resuls on Texure image Segmenaion In his aer fuzzy c means clusering used o segmening he exure image. The image is divided ino 32x32 overlaing blocks o segmen he image as ixel wise segmenaion. From each blocks NEW exure descrior is exraced o cluser he image using Fuzzy C-means clusering algorihm. To choose he K value auomaically, firs inu image convered ino gray scale and he hisogram is calculaed. This hisogram conains more number of unwaned eaks because of he some ercenage of noise and a considerable level of uncerainy. Hence, hisogram is smoohened using Gaussian filer and he number of eaks deeced. Fuzzy C-Means clusering algorihm clusers he exure image. Fig.5 (a) he original exure image; fig.5 (b) hisogram of he original image; fig.5 (c) he number of eaks deeced is four (K=4). So he segmened regions are four; segmened image is shown in fig.5 (d). Four differen exure images segmened 44

7 wih four differen colors. Blue color shows he exure 1 and red color shows he exure 2 and ink, green shows he exure 3 and exure 4. (e) Inu four exure image wih differen size (f) clusered region using Fuzzy C-Means algorihm. Same mehod alied ino K means clusering algorihm. K means clusering clusers he exure image.fig.6 (a) he original exure image; fig.6 (b) hisogram of he original image; fig.6 (c) he number of eaks deeced is four (K=4). So he segmened regions are four; segmened image is shown in fig.6(d). Four differen exure images segmened wih four differen colors. Blue color shows he exure 1 and green color shows he exure 2 and ink, red shows he exure 3 and exure 4. (e) Inu four exure image wih differen size (f) clusered region using K-Means algorihm Number of occurrence (a) Pixel Inensiy (b) Number of occurrence Pixel Inensiy (c) (d) (e) (f) Fig. 5. (a) Inu four exure image wih same size (b) hisogram of he image (c) Peaks deecion afer smoohing he hisogram (d) clusered regions using Fuzzy C-Means algorihm (e) Inu four exure image wih differen size (f) clusered region using Fuzzy C-Means algorihm 45

8 Number of occurrence (a) Pixel Inensiy (b) Number of occurrence Pixel Inensiy (c) (d) (e) (f) Fig. 6. (a) Inu four exure image wih same size (b) hisogram of he image (c) Peaks deecion afer smoohing he hisogram (d) clusered regions using K-Means algorihm (e) Inu four exure image wih differen size (f) clusered region using K-Means algorihm Fig.7 (a) shows he biolar lae samles, and same mehodology alied for his image. Green color shows he resence of carbon in he image, blue color shows he resence of eoxy in he image. 46

9 (a) (b) (c) (d) Fig. 7. (a) Biolar lae samles (b) Microscoic image (c) SEM image (d) clusered image using Fuzzy C-Means algorihm based on NEW exure descrior 2.6. Performance Analysis To es he erformance of he algorihm we have used hree measuremen arameers viz: Accuracy, Precision and Recall. These are calculaed from confusion marix Precision Precision is defined he raio of he number of relevan records rerieved o he oal number of irrelevan and relevan records rerieved. I is usually exressed as a ercenage. Pr ecision = + f (7) Where - rue osiive, n is rue negaive, f is false osiive and fn is false negaive Recall Recall is defined as he raio of he number of relevan records rerieved o he oal number of relevan records in he daabase. I is also exressed inerms of ercenage. Recall = + fn (8) 47

10 2.6.3 Accuracy Advanced Comuing: An Inernaional Journal (ACIJ), Vol.7, No.1/2, March 2016 Accuracy gives he measure of overall correcness of he roosed work, and is calculaed as he sum of correc clusers divided by he oal number of clusers. Accuracy = + n + + f n + f n (9) 3. CONCLUSION Table 1 feaure relevance Feaures Classifier Accuracy (%) NEWfeaure K-means 46% FCM 60.12% In his aer an effecive mehod o segmen he exure image based on exure feaure. In he roosed mehod, NEW exure feaure descrior is used o exrac he exure feaure. I allows he sysem o segmen he exure image using Fuzzy C-Means algorihm. For noiseless images, Fuzzy C-Means algorihm roduced he bes resuls comared o K-Means algorihm. Selecion of K value by auomaic sysem. Finally, exerimenal resuls are reored and invesigae he effeciveness of he roosed mehodology. REFERENCES [1] A.K. Jain and F. Farrokhnia (1991) UnsuervisedTexure Segmenaion Using Gabor Filers Paern Recogniion, vol.24, no. 12, [2] A.K.Jain and K. Karu(1996) Learning Texure Discriminaion Masks, IEEE Trans. Paern Analysis Machine Inelligence, vol. 18, no. 2, [3] Anil K. Jain (1998) Texure Analysis The Handbook of Paern Recogniion and Comuer Vision (2nd Ediion), [4] B. S. Manjunah, W. Y. Ma (2008) Texure Feaures for Browsing and Rerieval of Image Daa IEEE Transacions on Paern Analysis and Machine, Volume 18, Issue 8, Pages: [5] Cross,G.R., Jain, A.K ( 2007) Markov random field exure models IEEE Trans. Paern Anal. Machine Inell, PAMI-5(1), [6] C.S. Lu, P.C. Chung, and C.F. Chen(1997) Unsuervised Texure Segmenaion via Wavele Transform Paern Recogniion, vol. 30, no. 5, [7] Fernando Rodriguez, Guillermo Sairo (2008) Sarse Reresenaions for image classificaion: Learning discriminaive and reconsrucive non-arameric dicionaries IMA Prerin Series # [8] John Wrigh,y. yang(2009) Robus Face Recogniion via Sarse Reresenaion IEEE rans. On Paern Analysis and Machine Inelligence vol.31. [9] J. R. Smih and S.-F. Chang (2005) Visually searching he web for conen IEEE Mulimedia Magazine 4(3), [10] J. Yang, J.Wang, and T. S. Huang (2011) Learning he sarse reresenaion for classificaion In Proc. ICME, ages 1 6. [11] J. Yang, Z. Wang, Z. Lin,. Shu, and T. Huang (2012) Bilevel sarse coding for couled feaure sace. In Proc. CVPR. [12] J. Zhang and M. Marszalek (2006) Local Feaures and Kernels for Classificaion of Texure and Objec Caegories: A Comrehensive Sudy Inernaional Journal of Comuer Vision DOI: /s [13] M.Haralick, K. Shanmugam (1973) Texural feaures for Image Classificaion IEEE ansacions on sysem, man and cybernaics vol SMC-9,

11 [14] Mohamed Borchani, Georges Samon (2005) Texure feaures for Image Classificaion and Rerieval SPIE, Vol.3229, 401. [15] S.Z. Li (1995) Markov Random Field Modeling in Comuer Vision New York Sringer-Verlag. [16] Zhaowen Wang, Jianchao Yang(2013) A Max-Margin Persecive on Sarse Reresenaion-based Classificaion ICCV [17] S.M.M.Roomi, R.Raja, D.Dharmalakshmi, Classificaion and rerieval of naural scenes, Fourh Inernaional Conference on Comuing, Communicaions and Neworking Technologies (ICCCNT), DOI: /ICCCNT ,.1-8,2013. [18] Deeali Aneja and Tarun Kumar Rawa Fuzzy Clusering Algorihms for Effecive Medical Image Segmenaion, Inelligen Sysems and Alicaions, DOI: /ijisa ,-55-61, [19] R. Gonzalez and R. Woods Digial Image Processing, Addison-Wesley Publishing Comany, g 191, AUTHORS S.Mohamed Mansoor Roomi received his B.E.degree from Madurai Kamaraj Universiy, in 1990, his M.E. degree in Power Sysems and Communicaion Sysems from Thiagarajar College of Engineering in 1992 and 1997 and his Ph.D.in Image Analysis from Madurai Kamaraj Universiy in He has auhored and co-auhored more han 150 aers in various journals and conference roceedings and numerous echnical and indusrial rojec reors. M.Mareeswari received her B.Tech.degree from Thiagarajar College of Engineering, Madurai in G. Maragaham received her B.E.degree from Thiagarajar College of Engineering, Madurai, in 1983, her received M.E. degree from Shan College of Engineering, Tanjavur, in 2002.her received MS degree from BIT,Palani. 49