SAR Image Feature Extraction and Target Recognition Based on Contourlet and SVM

Transcription

1 00 3rd Internatona Conference on Computer and Eectrca Engneerng (ICCEE 00) IPCSIT vo. 53 (0) (0) IACSIT Press, Sngapore DOI: /IPCSIT.0.V53.o..79 SAR Image Feature Extracton and Target Recognton Based on Contouret and SVM A Wang, +, Chunhu Zhao Schoo of Informaton and Communcaton Engneerng, Harbn Engneerng Unversty Hgher Educatona Key Lab for Measurng & Contro Technoogy and Instrumentatons of Heongjang, Harbn Unversty of Scence and Technoogy, Harbn, Chna Abstract. In ths paper, we ntroduce a SAR mages target recognton approach based on the contouret transform and support vector machne (SVM), whch takes technoogca advantages of both SVM and the contouret transform for feature extracton. Keywords: Contouret; SVM; feature extracton. Introducton SAR systems are mportant sensors due to ther a weather, day/nght, ong standoff capabty. Aong wth the deveopment of radar technooges, as we as wth ncreasng demands for target recognton (ATR) usng SAR has become an actve research area. The process of SAR mage formaton of a target or a scene s a senstve functon of the range of the system and scene parameters. Hence, the ATR agorthms need to be extremey robust to any varaton n the magng system and scene parameters. Some of the reports wth exceent ATR performance reported have used a range of cassfcaton agorthm, startng from smpe tempate matchng and Gaussan-modeed Bayesan approach to those nvovng more nvoved agorthm such as the support vector machne approaches. The tempate-based method, however, s computatonay neffcent. Here, hgh cassfcaton accuracy condtons the creaton of arge sets of fters for each cass, whch resuts n extensve computaton both n the tranng and testng stages. As a key step of ATR, the feature extracton pays an mportant part n performng the recognton task. Many effcent approached for target feature extracton have been deveoped, e.g, Prncpe Component Anayss (PCA), Independent Component Anayss (ICA), Lnear Dscrmnant Anayss(LDA), Kerne PCA(KPCA), Kerne LDA(KLDA) and correspondng modfed methods [-7].. Contouret, proposed by Mnh.Do and Martn Vetter, s a rea method for -D representaton of mage. Contouret not ony process the man features of waveets n mut-resouton anayss, but aso can specay decompose the subbands at each scae nto dfferent parts wth fexbe number. The Contouret Transform (CT) was chosen as the bass for feature extracton methodoogy proposed. By anayzng each target n dfferent scaes t s possbe to see characterstcs that were not obvous n the orgna resouton [8-0]. The paper s organzed as foows. In Secton II, we present the prncpe of the contouret transform. In Secton III, we gve a bref revew of SVM. Secton IV descrbes deta steps of our target recognton approach, ncudng feature extracton and cassfcaton. Expermenta resuts on the MSTAR data are presented and dscussed n Secton V. We concude wth our remarks on the proposed agorthm n Secton VI. + Correspondng author. E-ma address: a95@hrbust.edu.cn.

2 . Contouret Transform The contouret transform provdes a mutresouton, oca and drectona mage expanson. It many has two stages: mutscae Lapacan Pyramd decomposton and the drectona fter bank. As shown n Fgure, t has a doube fter bank structure and the two stages are ndependent each other. The drectona fter bank has a fexbe number of drectons and t ony capture the hgh frequency of the nput mage because the ow frequences of the nput mage are removed before appyng t. (,) bandpass drectona subbands mage bandpass drectona subbands Fgure. The bock dagram of contouret transform The decomposton coeffcents of contouret are smar wth waveet coeffcents n that most coeffcents are amost zero and ony few of them, near the edge of the objects n the mage, have arge magntudes. Besdes ths feature, the characterstc of coeffcents have strong dependences across decomposton eves. The dfferent drectons n each decomposton eve can descrbe dfferent features for same edge. Contouret s a geometry of transformaton based on mage. The mut-scae decomposton and the drectona decomposton are two ndependent processes whch w effectvey to express the contours and texture-rch of mage. They have ong strp structure whch aspect ratos have changed wth the scae n the enongated supported, t can effectvey to track the characterstcs of near dscontnutes and area dscontnutes n the mage. Compared wth waveet, contouret has a rch basc functon whch can descrbe the smooth edges usng ess transform coeffcents, and can take the pont dscontnutes whch have the same drecton together nto a near or area dscontnutes. 3. SVM Cassfer Based on structura-rsk mnmzaton prncpes, SVM has acheved remarkabe success n patternrecognton probems. By non-near mappng, t projects sampes from ow-dmenson space to hghdmenson space and mpements the non-near cassfabe probem n the hgh-dmenson. SVM aways fnds a goba mnmum because t usuay tres to mnmze a bound on the structura rsk, rather than the emprca rsk. Emprca rsk s defned as measured mean error rate on the tranng set as beow Remp ( α ) = y f ( x, α ) () where s number of observaton, y s cass abe and x s sampe vector. And structura rsk s defned as a structure of dvded entre cass of functon nto nested subset and fndng the subset of functon whch mnmzes the bound on the actua rsk. SVM acheves ths goa by mnmzng the foowng Lagrangan formuaton: L p = w ( xw + b) + α y α () where α s postve Lagrange mutper. When deang wth nonnear probem, we use the kerne trck to maxmum-margn hyperpanes. In ths paper, we use the Gaussan Rada bass functon

3 x x j k ( x, x j ) = exp σ (3) Essentay, object recognton s a compex mut-cassfcaton probem. However, SVM s orgnay suted for two-cass probem. When deang wth mut-cassfcaton probem, t s abe to construct a mutcass cassfer by combnng severa bnary cassfers. 4. Proposed Recognton Method In the proposed agorthm, we ntend to fnd the best target representaton whe keepng the dmenson of the feature vector reasonaby ow [-]. The feature of the owpass and bandpass drectona subbands are extracted respectvey [4,5]. Energy, entropy of deta coeffcents, energy of approxmaton, mean, varance, skewness are seected as dfferent contouret doman features. For an subband c (, j), energy s cacuated as: Entropy s defned as: (, j) c j= E = (4) H = M k = p k og p k (5) where M s the number of random varabe, α s the random varabes wth probabtes p. Mean and varance are defned as: Mean = Var = c, j ( j) [ c( j) Mean], j Shewness s a measure of the asymmetry of the data around the sampe mean. The skewness of a dstrbuton s defned as: ( c Mean) E S = (8) 3 Var E represents the expected vaue of the quaty c. Then the features are merged and the SVM s used for target recognton n SAR mages. where () t 5. Experments and Performance Evauatons The SAR mage dataset used for expermentaton was coected under the Movng and Statonary target Acquston and Recognton (MSTAR) program. It s composed of X band SAR mages for 5 vehce targets taken at two depresson anges, 5 and 7 degrees, and dfferent target orentaton between 0 and 360 degrees. The resouton s ft by ft. Images of depresson ange 7 degrees were used for 7 degrees were used for constructng the tranng set, whereas those of depresson ange 5 degrees were used for testng. Unke optca mages, SAR mages refect the structure of target s pont scatters, whch does not have even refect rate over dfferent orentaton anges. Thus SAR mages of the same target taken at dfferent orentaton anges can show great dfferences, whch make cassfcaton even more dffcut [6-9]. Imagery dstrbuton s detaed n Tabe I, whereas sampes of these mages are shown n Fg.. The vehces n MSTAR dataset contan the BMP (S_9563, S_9566, S_C) tracked Armored Personne Carrer, the BTR70 (S_C7) wheeed Armored Personne Carrer and the T7 (S_3, S_8. S_S7) Man Batte Tank. Dfferent sera numbers n one-target cass mean that vehces are varants wth sma dfferences n confguraton, artcuaton and damage under extended operatng condtons, e.g., tanks wth (6) (7)

4 drum mounted or unmounted, carrers wth antenna extended or shunk [3]. The proposed CT-SVM method can obtan 7% hgher recognton rate than waveet-based SVM. It was shown that the accuracy of the contouret transform especay n remote sensng data acquston systems. Tabe I Summary of MSTAR Database Type Sera umber Tranng Set Sze Testng Set Sze S_ BMP S_ S_C BTR70 S_C T7 S_ S_ S_S7 8 9 (a) BMP (b) BTR70 (c) T7 Fgure. Sampes of SAR target 6. Concuson SVM motvated by the statstca earnng theory s a probabstc non-near mode. Snce theoretcay attractve features and profound emprca performance, SVM s ganed popuarty n many areas. Ths paper proposes a new approach for target recognton n SAR mages based on expotng the features of the contouret transform combned wth SVM cassfcaton. In partcuar, we have anayzed ths technque both theoretcay and expermentay. The effcency of our method has been demonstrated on standard databases MSTAR. The experments resuts show that the features are effectve and robust to targets wth the same confguraton of confguraton dfferences. 7. Acknowedgment Ths work was supported n part by the Chna Postdoctora Scence Foundaton ( ). 8. References [] L Maokuan, Guan Jan, Huan Hu and Gao Xn, SAR ATR based on support vector machnes and ndependent component anayss, CIE Internatona Conference of Radar Proceedngs, 006, October, Shangha, pp [] Huan Ruohong, Yang Ruang, SAR mage target recognton based on KFD+ICA feature extracton, Systems Engneerng and Eectroncs, 30(7), Juy 008, pp [3] Yjun Sun, Zhpeng Lu, Snsa Todorovc and Jan L, Adaptve boostng for SAR automatc target recognton, IEEE Transactons on Aerospace and Eectronc Systems, January 007, 43(), pp.-5. [4] A.K. Mshra and B. Mugrew, Bstatc SAR ATR, IET Radar Sonar avng, 007, (6): [5] Joseph A.O suvan, Mchae D.Devore, Vkas Keda and Mchae I.Mer, SAR ATR performance usng a condtonay Gaussan mode, IEEE Transactons on Aerospace and Eectronc Systems, Jan. 00, 37():9-08.

5 [6] Sef E-Dawaty, Hossam Osman and Hussen I. Shahem, Enhanced SVM versus severa approaches n SAR target recognton, Internatona Conference on Computer Engneerng and Systems, ov. 006, pp [7] Qun Zhao, Prncpe J C, Support Vector Machne for SAR automatc target recognton, IEEE Transacton Aerospace and Eectronc Systems, 00, 37(), pp [8] Do.M.. and Vetter.M, The contouret transform: an effcent drectona mutresouton mage representaton, IEEE Transactons on Image Processng, Dec. 005, 4(), pp [9] Po.D.D.-Y and Do. M., Drectona mutscae modeng of mages usng the contouret transform, IEEE Transactons on Image Processng, June 006, 5(6), pp [0] Arthur L. Cunha, Japng Zhou and Do M., The nonsubsamped contouret transform: theory, desgn, and appcatons, IEEE Transactons on Image Processng, Oct. 006, 5(0), pp [] Yang Shuyuan, Wang Mn and Jao Lcheng, Radar target recognton usng contouret packet transform and neura network approach, Sgna Processng, Apr. 008, 89(4), pp [] Ru Hu, Lcheng Jao, Weda Zhou and Y Gao, Remote sensng target recognton based on contouret and kerne fsher dscrmnant, Internatona Conference on Computaton Integence and Securty, 007, pp [3] Douve P.L. Measured and predcted Synthetc Aperture Radar target comparson, IEEE Transactons on Aerospace and Eectronc Systems, 38(), pp.5-37, 00 [4] Han Zheng, Su Zhgang, Han Png and Wu Renbao, SAR target recognton method based on orthogona subspace of sampes, Journa of Eectroncs & Informaton Technoogy, 3(), 009, pp [5] Wang Shx and He Zhguo, The fast target recognton approach based on PCA features for SAR mages, Journa of atona Unversty of Defense Technoogy, 008, 30(3), pp [6] Hu Yng, Wang Shuang, Hou Bao and Jao Lcheng, Remote sensng target recognton based on SWBCT and projecton feature, Journa of Infrared and Mmeter Waves, 007, 6(6), pp [7] Hu Lpng, Lu Jn and Lu Hongwe, et a., Automatc target recognton based on SAR mages and two-stage DPCA features, Journa of Eectroncs & Informaton Technoogy, 008, 30(7), pp [8] Yn Chen, Genshe Chen and Rck S. Bum, et a., Image quaty measures for predctng automatc target recognton performance, IEEE Aerospace Conference, 008 Mar.,pp.-9. [9] Reza Javdan, M.A. Masnad-Shraz, Z.Azmfar and M.H.Sadreddn, A comparatve study between waveet and contouret transform features for textura mage cassfcaton