Comparitive Analysis of Image Segmentation Techniques

Transcription

1 ISSN: Volume, Issue 9, Sepember 3 Compariive Analysis of Image Segmenaion echniques Rohi Sardana Pursuing Maser of echnology (Compuer Science and Engineering) GJU S& Hissar, Haryana Absrac Image segmenaion is he process of pariioning an image ino muliple segmens, so as o change he represenaion of an image ino somehing ha is more meaningful and easier o analyze. Several general-purpose algorihms and echniques have been developed for image segmenaion. In his paper, we presen osu mehod, waershed mehod and Color- ased Segmenaion Using K-Means Clusering for image segmenaion. hen evaluaion of hese mehod is done using four evaluaion merics: probabilisic Rand index, global consisency error, variaion of informaion and peak signal o noise raio. We inend o find ou he bes algorihm using evaluaion merices. Keywords: Image segmenaion, hreshold, Osu mehod, Waershed, Color-ased Segmenaion Using K-Means Clusering, PRI, GCE, VOI, PSNR,. Inroducion Image Segmenaion is a common process in an image analysis especially in he field of vision and racking. Segmenaion is defined as a mehod ha subdivides an image ino is consiuen regions or objecs. he level o which he subdivision is carried depends on he problem being solved. ha is, segmenaion should sop when he objec of ineres in an applicaion have been isolaed []. Mahemaical Form Mahemaically if he domain of image is given by I, hen he segmenaion problem is o deermine he ses S j, whose union is enire Image I. hus he ses ha make up segmenaion mus saisfy I = n () Sj j= where Sj Sk = φ for k jand each S j is conneced and n is number of objecs of ineres. Image Segmenaion echniques image segmenaion is called he hresholding mehod. his mehod is based on a hreshold value o urn a gray-scale image ino a binary image. Anoher image segmenaion mehod is Edge based Mehod ha is more common for deecing disconinuiies in gray level han deecing isolaed poins and hin lines because isolaed poins and hin lines so no occur frequenly in mos pracical images[]. Anoher mehod Graph ased Segmenaion is a fas and efficien mehod of generaing a se of segmens from an image.he graph based image segmenaion is based on selecing edges from a graph, where each pixel corresponds o a node in he graph[3].in his paper, Osu hresholding Algorihm, Waershed Algorihm and Color-ased Segmenaion Using K-Means Clusering are sudied. Comparison of hese algorihm are done using performance merics. Prediced daase is compared wih ground ruh daa. Mehodology. Osu s hresholding Mehod Osu [4] proposed a dynamic hresholding selecion mehod in 979. his mehod suggess maximizing he weighed sum of beween-class variances of foreground and background pixels o esablish an opimum hreshold. Osu s hresholding echnique is based on a discriminae analysis which pariions he image ino wo classes C and C a gray level such ha C = {,,3,.,} and C = { +,+,.,L-}, where L is he oal number of he gray levels of he image. Le he number of pixels a he ih gray level be n i and n be he oal number of pixels in a given image. he probabiliy of occurrence of gray level i is defined as: pi = ni n () Many algorihms and mehods have been developed for image segmenaion. he simples mehod of 65

2 ISSN: Volume, Issue 9, Sepember 3 C and C are normally corresponding o he objec of ineresed and he background, he probabiliies of he where wo ypes of fairly disinc classes exis in he image [5]. wo classes are ω and ω : pi (3) i L i pi (4) hus, he means of he wo classes can be compued as: ( ) μ i ipi ( ) ( ) (5) L ipi () (6) ω () i Le σ and σ be he beween-class variance and oal variance respecively. An opimal hreshold can be obained by maximizing he beween-class variance. Arg max il (7) Where, he beween-class variance σ and σ are defined as: σ =ω( μ -μ ) +ω( μ -μ ) (8) L i ( i ) (9) he oal mean of he whole image μ is defined as: An equivalen hreshold = Arg μ L = ip i i= () formula for obaining opimal is as follows: Max { ωo( μ -μ ) +ω( μ -μ ) } L () Osu s mehod of hresholding gray level images is efficien for separaing an image ino wo classes. Waershed Algorihm he waershed ransform finds cachmens basins and waershed ridge lines in an image by reaing i as a surface where ligh pixels are high and dark pixels are low. One of he mos imporan drawback associaed o he waershed ransform is he over segmenaion ha commonly resuls. he usual way of predeermining he number and approximae locaion of he regions provided by he waersheds echnique consiss in he modificaion of he homoopy of he funcion o which he algorihm is applied. his modificaion is carried ou via a mahemaical morphology operaion, geodesic reconsrucion [6], by which he funcion is modified so ha he minima can be imposed by an exernal funcion (he marker funcion). All he cachmen basins ha have no been marked are filled by he morphological reconsrucion and so ransformed ino non minima plaeaus, which will no produce disinc regions when he final waersheds are calculaed. Segmenaion using he waershed ransform works well if you can idenify, or mark, foreground objecs and background locaions [7]..3 Color-ased Segmenaion Using K-Means Clusering Color-ased Segmenaion using K-Means follows he following seps:-.read he color image..conver image from RG color space o LA color space. 3.Classify he colors in A space using K-Means Clusering. 4.Label every pixel in he image using he resuls from KMeans. 5.Creae Images ha segmen he image by color. 3 Performance Merics For evaluaing he performance of segmened image, we use following merics. 3. Probabilisic Rand Index (PRI) Rand Index is he funcion ha convers he problem of comparing wo pariions wih possibly differing 66

3 ISSN: Volume, Issue 9, Sepember 3 number of classes ino a problem of compuing pair wise label relaionships. PRI couns he fracion of pairs of pixels whose labelling are consisen beween he compued segmenaion and he ground ruh, averaging across muliple ground ruh segmenaions o accoun for scale variaion in human percepion. I is a measure ha combines he desirable saisical properies of he Rand index wih he abiliy o accommodae refinemens appropriaely. Since he laer propery is relevan primarily when quanifying consisency of image segmenaion resuls. Consider a se of manually segmened (ground ruh) images {S, S,..., S K } corresponding o an image X = {x, x,... x i,..., x N }, where a subscrip indexes one of N pixels. Ses is he segmenaion of a es image, and hen PRI is defined as: PR( S,{ Sk }) = N [ c es ij ij + ( - i,j i< j p c ij )(- p ij) ] () Here cij denoe he even of a pair of pixels i and j having he same label in he es image S es : S cij = I(l = l i es Ses j ) (3) his measure akes values in [, ] when S es and {S, S,..., S K } have no similariies and when all segmenaions are idenical[8]. 3. Global Consisency Error (GCE) he Global Consisency Error (GCE) measures he exen o which one segmenaion can be viewed as a refinemen of he oher [9].I is a Region-based Segmenaion Consisency, which measures o quanify he consisency beween image segmenaions of differing granulariies. I is used o compare he resuls of algorihms o a daabase of manually segmened images. Le S and S be wo segmenaion as before. For a given poin x i (pixel), consider he classes (segmens) ha conain x i in S and S. hese ses are denoed in he form of pixels by C (S, x i ) and C (S, x i ) respecively []. GCE (S min{ i x( S,S ), i x( S,S ) = i i,s )} (4) n 3.3 Variaion of Informaion (VOI) I measures he sum of informaion loss and informaion gain beween he wo class, and hus i roughly measures he exen o which one class can explain he oher. he VOI meric is nonnegaive, wih lower values indicaing greaer similariy. I is based on relaionship beween a poin and is class. I uses muual informaion meric and enropy o approximae he disance beween wo classes across he laice of possible classes. More precisely, i measures he amoun of informaion ha is los or gained in changing from one class o anoher (and, hus, can be viewed as represening he amoun of randomness in one segmenaion which canno be explained by he oher). he variaion of informaion is a measure of he disance beween wo class (pariions of elemens). A class wih pixels X,X,,,,,X k is represened by a random variable X wih X={.K} such ha p i = X i /n iєx and n= i X i he variaion of informaion beween wo class X and Y so represened is defined o be VI( X,Y ) = H( X ) + H(Y )-I( X,Y ) (5) where H(X) is enropy of X and I(X,Y) is muual informaion beween X and Y. VI(X,Y) measures how much he pixel assignmen for an iem class X reduces he uncerainy abou he iem's pixel in class Y []. 3.4 Peak signal o noise raio (PSNR) PSNR is used o measure he difference beween wo images. I is defined as PSNR = log(b/rms) where b is he larges possible value of he signal (ypically 55 or ), and rms is he roo mean square difference beween wo images. he PSNR is given in decibel unis (d), which measure he raio of he peak signal and he difference beween wo images[]. 4 Experimenal evaluaion For Segmenaion, Original images and Image Mask are aken form erkeley Daabase. Image segmenaion is done by using hree echniques: () Osu Mehod () Waershed Mehod 67

4 ISSN: Volume, Issue 9, Sepember 3 (3) Color-ased Segmenaion Using K-Means Clusering Experimen 4. : Osu Mehod. Read image wih gray levels of =[, L]. Compue hisogram and probabiliies of each inensiy level ω i( ) = μ i( ) = 3. Se up iniial and 4. Sep hrough all possible hresholds =[, L] maximum inensiy ω μ o Compue i and i o Compue σ ( ) 5. Desired hreshold corresponds o he maximum σ ( ). Figure 4. : Waershed Image Segmenaion Experimen 4.3 : Color- ased Segmenaion Using K-Means Clusering. Read he color image.. Conver image from RG color space o LA color space. 3. Classify he colors in A space using K- Means Clusering. 4. Label every pixel in he image using he resuls from KMeans. Figure 4. : Osu Image Segmenaion 5. Creae Images ha segmen he image by color. Experimen 4. : Waershed Mehod. Read he color image and conver i o gray scale.. Use he gradien magniude as segmenaion funcion. 3. Mark he foreground objecs. 4. Compue he background markers. 5. Compue he waershed ransform of he segmened funcion. 6. Visualize he resul. Figure 4.3 : Color-ased Segmenaion Using K-Means Clusering Experimen 4.4 : Ground ruh For Ground ruh, we superimposed Image Mask on Original Image. 68

5 ISSN: Volume, Issue 9, Sepember 3 MERICS OSU WAERSHED Color-ased Segmenaion Using K-Means Clusering PRI GCE Figure 4.4 : Mask Image Superimposed On Original Image Experimen 4.5 : Differen Images Experimens Perform on We perform experimen on differen images from erkeley daase. One of he image is considered in his paper. Now we have o find which segmenaion algorihm is bes. For his we ake image and image mask from erkeley Daabase. Ground ruh is obained by superimposed he image mask on original image. Osu Image, Waershed Image and Color-ased Segmenaion using K-Means Clusering is resul as shown in fig. VOI PSNR able 4. : Comparison Using Parameer PRI, GCE, VOI,PSNR References [] Linda G. Shapiro and George C. Sockman Compuer Vision, Upper Saddle River, New Jersey: Prenice Hall, pp ,. [] Salem Saleh Al-amri, Dr. N.V. Kalyankar and Dr. Khamikar S.D Image Segmenaion by Using Edge Deecion Inernaional Journal on Compuer Science and Engineering,Vol., No. 3,pp ,. [3] Sandeep Chalasani Graph ased Image Segmenaion [4] N. Osu, A hreshold selecion mehod from gray-level hisogram, IEEE ransacions on Sysems Man Cyberne, pp. 6-66, 978. [5] WANG Hongzhi, DONG Ying An Improved Image Segmenaion Algorihm ased on Osu Mehod Inernaional Symposium on Phooelecronic Deecion and Imaging 7: Relaed echnologies and Applicaions, Vol. 665,8. [6] Ashwin Kumar, Pradeep Kumar A New Framework for Color Image Segmenaion Using Waershed Algorihm Journal of Elecronic Imaging,Vol.,No. 3,pp. 4-46,. Figure 4.5 : OSU, WAERSHED & COLOR-ASED SEGEMENAION USING K-MEANS CLUSERING able 4. shows he PRI, GCE, VOI & PSNR of Osu Image, Waershed Image & Color- ased Segmenaion Using K-Means Clusering Image of Image. his shows PRI, GCE of Osu Image is higher han he oher mehods and VOI of Osu Image is low as compare o oher mehods. So using PRI, GCE, VOI, PSNR we conclude ha Osu Mehod is beer han oher mehods. [7] Mandeep Kaur, Gagandeep Jindal Medical Image Segmenaion using Marker Conrolled Waershed ransformaion IJCS Vol., Issue 4,. [8] R. Unnikrishnan C. Panofaru M. Heber, A Measure for Objecive Evaluaion of Image Segmenaion Algorihms Proceedings of IEEE Compuer Sociey Conference on Compuer Vision and Paern Recogniion,Vol. 3,Page 34,5. [9] Allan Hanbury, Julian Soinger, On segmenaion evaluaion merics and region coun. [] Manisha Sharma, Vandana Chouhan Objecive Evaluaion Parameers of Image Segmenaion Algorihms Inernaional Journal of Engineering and Advanced echnology (IJEA) Vol., Issue,. 69