ISSN(Onlne): 2320-9801 ISSN (Prnt): 2320-9798 Internatonal Journal of Innovatve Research n Computer and Communcaton Engneerng (An ISO 3297: 2007 Certfed Organzaton) Vol.2, Specal Issue 1, March 2014 Proceedngs of Internatonal Conference On Global Innovatons In Computng Technology (ICGICT 14) Organzed by Department of CSE, JayShrram Group of Insttutons, Trupur, Tamlnadu, Inda on 6 th & 7 th March 2014 Effcent Segmentaton and Classfcaton of Remote Sensng Image Usng Local Self Smlarty Karthkeyan.K 1, Thruselv.M 2 M.E (Appled Electroncs), SNS College of Engneerng, Combatore, Tamlnadu, Inda 1, 2 ABSTRACT- Segmentaton and classfcaton are mportant role n remote sensng mage analyss. Recent research shows wth the am of mages can be descrbed n herarchcal structure or regons. In ths proect, we submt applcaton graph laplacan energy as generc measure for segmentaton. We capture n geometrc outlne of regon n an mage by usng apply local self smlarty features. Ths paper fnds applcaton n remote sensng mage analyss. It decreases the redundancy n the herarchy by order of magntude wth small loss or performances. We have acheved better performance from graph laplacan energy method. I mprove the effcency usng unsupervsed learnng. I. INTRODUCTION In ths paper, we suggest a new herarchcal mage analyss method that apples the graph Laplacan energy (LE) [9] as a generc measure for segmentaton. Wth segmentaton results avalable, we contnue to the classfcaton step usng local self-smlarty (LSS) [10] to ntegrate the local contextual and shape nformaton. We exhbt the effectveness of the proposed classfcaton method n urban-area land-cover classfcaton usng VHR remote sensng mages. The contrbutons of ths paper are threefold. Frst, we use the graph LE to explan the VHR remote sensng mage n a herarchcal structure. It places of nterest the hgh-level semantc structure of an mage. Second, we ntroduce a method to extort local contextual and shape nformaton n local regons usng LSS. Fnally, we exhbt the feasblty of a classfcaton system for remote sensng mages by combne both hgh-level and low-level nformaton. Ths paper s organzed as follows. Secton II explans the segmentaton method usng the graph LE. Secton III descrbes the classfcaton step by extractng LSS features and tranng support vector machnes (SVMs). Fnally, n Secton IV, we present the expermental results, whch nclude comparsons between the proposed method and dfferent classfcaton approaches n the lterature. Conclusons of ths paper are drawn n Secton V. II. HIERARCHICAL REPRESENTATION USING GRAPH LE Fgure. 1. Example of watershed segmentaton. The left hand s the orgnal mage. The rght hand shows the correspondng watershed segmentaton results. Copyrght @ IJIRCCE www.rcce.com 2329
ISSN(Onlne): 2320-9801 ISSN (Prnt): 2320-9798 Internatonal Journal of Innovatve Research n Computer and Communcaton Engneerng (An ISO 3297: 2007 Certfed Organzaton) Vol.2, Specal Issue 1, March 2014 Proceedngs of Internatonal Conference On Global Innovatons In Computng Technology (ICGICT 14) Organzed by Department of CSE, JayShrram Group of Insttutons, Trupur, Tamlnadu, Inda on 6 th & 7 th March 2014 Let G be an (n,m) graph wth n vertces and m edges and A be ts adacency matrx. Let d be the degree of the th vertex of G and D be the correspondng degree matrx, where D (, ) = d. Then, L = D A s the Laplacan, and the graph LE s defned as LE(G) n 1 2m n..(1) Where denotes egen values of the Laplacan matrx and 2m/n s the average vertex degree. Based on ths defnton, we propose a novel method to take out herarchcal structure from remote sensng mages. Frst, we compute the mage gradent and cut the remote sensng mages nto small connected regons usng watershed segmentaton [8]. An example of the segmentaton results s shown n Fgure. 1. Ths tread generates the bottom level of the herarchy and transforms the nput mage nto a regon adacency graph. The subsequently step s to merge neghborng regons from bottom up pattern to create a herarchcal tree descrpton. In each mergng teraton, we merge the most related pars of neghborng regons and treat the newly merged regons as parent nodes. The mergng contnues awatng there s only one regon left. The ntal watershed segmentaton consequences are set as level 1. The mergng can be handlng all arbtrary shape. Each regon measured as sngle Gaussan = ( N,,C), n whch s a regon, N ndcates the number of pxels n ths regon, and μ and C are the mean and covarance of the property vectors at every pxel, respectvely. If two regons ( and ) can be merged, the modern regon ( = + ) s represented as NC C e N,, N...(2) The error term e compensate the egen space for the dfferentaton between the means of the model, whch s NN T e(, ) ( ) ( ) N N.(3) To estmate the graph LE of each merged regon, we need to get the adacency matrx at each mergng step. The adacency matrx s based usng the error measure. If regons and are lnked, the weght s w (, ) = exp( e(, ) / e max ), n whch e e max s the maxmum value of e(, ) above all connected pars. In every one mergng teraton, we merge the most related par of neghborng regons,.e., the par wth the smallest value of e(, ). Ths mergng step generates a complete herarchy tree. In the next step, we consder ths tree va graph LE. Our prncple s to select those tree levels that are lower n complexty than ther adacent levels. Copyrght @ IJIRCCE www.rcce.com 2330
ISSN(Onlne): 2320-9801 ISSN (Prnt): 2320-9798 Internatonal Journal of Innovatve Research n Computer and Communcaton Engneerng (An ISO 3297: 2007 Certfed Organzaton) Vol.2, Specal Issue 1, March 2014 Proceedngs of Internatonal Conference On Global Innovatons In Computng Technology (ICGICT 14) Organzed by Department of CSE, JayShrram Group of Insttutons, Trupur, Tamlnadu, Inda on 6 th & 7 th March 2014 Note that the standard graph LE n (1) s generally defned on a sngle connected graph. However, n our case, we take care of the merged regons as connected graphs so that each prmary watershed regon s a node n a graph. At each level except the maxmum one, there are added than one connected graphs. Therefore, we extend the orgnal graph LE defnton to make t proper for computng LE at each level of the tree. We establsh the normalzed graph LE (ngle). For a level wth K connected graphs, we defne the E= ().(4) Where G s the th connected graph of n nodes and n s the amount of nodes of all the assocated graphs. In the case of a sngle connected graph, our crcle of phrase reduces to the orgnal graph energy. We calculate the ngle at each level n the herarchy ndependently and use the ngle as a functon of level ndex. Fgure. 2 shows a typcal ngle curve. In ths curve, local mnma are met when graphs at partcular levels demonstrate homogeneous node degree, whch way that the graphs are close to regular graphs. They are n contact to levels that are less complex compared to the adacent levels. We desre the hghest level partton that gves the local mnmum, whch breaks the mage nto the least large components. Fgure.2. Graph energy as a functon of level ndex.the plot s computed usng (4) on the mage n fg.1. Copyrght @ IJIRCCE www.rcce.com 2331
ISSN(Onlne): 2320-9801 ISSN (Prnt): 2320-9798 Internatonal Journal of Innovatve Research n Computer and Communcaton Engneerng (An ISO 3297: 2007 Certfed Organzaton) Vol.2, Specal Issue 1, March 2014 Proceedngs of Internatonal Conference On Global Innovatons In Computng Technology (ICGICT 14) Organzed by Department of CSE, JayShrram Group of Insttutons, Trupur, Tamlnadu, Inda on 6 th & 7 th March 2014 Fgure.3. segmentaton process. Two levels of dvson, as shown n Fgure. 3. The mergng and segmentaton processes are shown n Fgure. 3. In ths example, the hgh-resoluton remote sensng mage has been segmented n a herarchcal manner. From ths procedure, a set of parts for each mage can be created, as shown at the bottom row n Fgure. 3. III. CLASSIFICATION USING LSS Fgure.4. process of extractng the LSS descrptor of pxel p. (a) Image patch and ts close neghborhood regon. (b) local nternal correlaton surface.(c) bnned log-polar representaton.(d) normalzed log-polar vector. The herarchcal descrpton of the mage forms the begnnng to the classfcaton step. To extract features from every one segmented parts, the LSS method [10] s used. The LSS descrbes the comparson between a patch and ts neghborng regon n an mage. It offers a sngle unfed way to descrbe the nsde relatons n an mage. The method of extractng the LSS descrptor s shown n Fgure.4. It s computed as follows. Copyrght @ IJIRCCE www.rcce.com 2332
ISSN(Onlne): 2320-9801 ISSN (Prnt): 2320-9798 Internatonal Journal of Innovatve Research n Computer and Communcaton Engneerng (An ISO 3297: 2007 Certfed Organzaton) Vol.2, Specal Issue 1, March 2014 Proceedngs of Internatonal Conference On Global Innovatons In Computng Technology (ICGICT 14) Organzed by Department of CSE, JayShrram Group of Insttutons, Trupur, Tamlnadu, Inda on 6 th & 7 th March 2014 1) Determne an N N correlaton surface ζp of an ω ω patch tp wth the mmedate N N regon Rp usng sum of squared dfferences (SSD) method. In ths letter, N = 30 and ω = 5. Both Rp and tp are centered at p. ζp(x) s the correlaton of tp wth an ω ω patch tx centered at x : ζp(x) = exp( (SSD(tp, tx)/δ)), where δ represents the maxmal varance of the varaton between every part of patches wthn a very tny neghborhood of p and the patch centered at p. 2) Dscretze the correlaton surface ζp on a log-polar grd, and accumulate the maxmal value of ζp wthn each grd bn dp (p, d) = maxx BIN (p,d){ζp(x)}. 3) Destablze the bnned log-polar vector by lnearly stretchng ts deals to the range [0, 1]. From each one mage part generated n Secton II, a set of LSS descrptors can be generated. In order to explan each part usng a sngle vector, the bag-of-vsual-words model [6] s adopted. We had t k-means clusterng method for vsual word encodng (k = 300 n our experments), whch groups the LSS descrptors nto k clusters. The cluster centers are clear as the vsual words, and a vsual vocabulary s constructed to llustrate the content of obects. After handng over each descrptor to the closest vsual word, all mage parts can be represented as a hstogram by plus the occurrence numbers of the vsual words. In the classfcaton step, the SVM s worn as the classfer. The selecton s for the most part based on the truth that t s one of the state of- the-art classfers on bag-of-vsual-words mage representaton. We mplement the C-SVC n LIBSVM [3] wth an RBF kernel. The parameters for the SVM are obtaned wth cross valdaton on a subset of physcally labeled tranng parts. Fgure.5. Example on classfcaton result (a) experment area. (b) Reference map. (c) Classfcaton result. Copyrght @ IJIRCCE www.rcce.com 2333
ISSN(Onlne): 2320-9801 ISSN (Prnt): 2320-9798 Internatonal Journal of Innovatve Research n Computer and Communcaton Engneerng (An ISO 3297: 2007 Certfed Organzaton) Vol.2, Specal Issue 1, March 2014 Proceedngs of Internatonal Conference On Global Innovatons In Computng Technology (ICGICT 14) Organzed by Department of CSE, JayShrram Group of Insttutons, Trupur, Tamlnadu, Inda on 6 th & 7 th March 2014 IV. CONCLUSION In ths paper, we have ntroduced graph LE as a dffculty calculate for remote sensng mage analyss. Wth the chosen herarches usng ths power, we can get a levelheaded semantc explanaton n terms of obects and obect parts whch help to acheve more robust classfcaton. We also ntroduced the LSS for urban-area land-cover classfcaton n remote sensng mages. The planned method has acheved performance on satellte mage analyss that s better than those from substtute methods. In the future, we plan to further explore the sem supervsed knowledge methods n satellte mage analyss. Ths allows the one after the other n unlabeled data to be extracted and used. Furthermore, tradtonal spectral features wll be ncorporated nto our system. I mprove the effcency usng unsupervsed learnng. REFERENCES 1. Akçay H. G. and Aksoy S., Automatc detecton of geospatal obects usng multple herarchcal segmentatons, IEEE Trans. Geosc. Remote Sens., vol. 46, no. 7, pp. 2097 2111, Jul. 2008. 2. Blaschke T., Lang S., and Hay G. J, Obect-Based Image Analyss. New York: Sprnger-Verlag, 2008. 3. Chang C.-C. and Ln C.-J. (2011, Apr.). LIBSVM: A lbrary for support vector machnes. ACM Trans. Intell. Syst. Technol. [Onlne]. 2(3), pp. 27:1 27:27. Avalable: http://w ntu.edu.tw/cln/lbsvm ww.cse. 4. Comancu D. and Meer P., Mean shft: A robust approach toward feature space analyss, IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 5, pp. 603 619, May 2002. 6. Fe-Fe L., Fergus R., and Torralba A., Recognzng and learnng obect categores, n Proc. CVPR Short Course, 2007 7. Gaetano R., Scarpa G., and Pogg G., Herarchcal texture-based segmentaton of multresoluton remote-sensng mages, IEEE Trans. Geosc. Remote Sens., vol. 47, no. 7, pp. 2129 2141, Jul. 2009. 8. Gauch J. M., Image segmentaton and analyss va multscale gradent watershed herarches, IEEE Trans. Image Process., vol. 8, no. 1, pp. 69 79, Jan. 1999. 9. Gutman I. and Zhou B., Laplacan energy of a graph, Lnear Algebra Appl., vol. 414, no. 1, pp. 29 37, Apr. 2006. 10. E. Shechtman and M. Iran, Matchng local self-smlartes across mages and vdeos, n Proc. IEEE CVPR, Jun. 17 22, 2007, pp. 1 8. Copyrght @ IJIRCCE www.rcce.com 2334