Remote Sensing Image Sequence Segmentation Based on the Modified Fuzzy C-means

Transcription

1 28 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY 2009 Reote Sensng Ige Sequence Segentton Bsed on the Modfed Fuzzy -ens Du Gen-yun Key Lb of Erth Explorton nd Inforton Technques of Educton Mnstry of hn, hengdu, Schun 60059, hn; ollege of oputer Scence nd Technology, Xuchng Unversty, Xuchng Henn 46000, hn; ollege of Inforton Engneerng, hengdu Unversty of Technology, hengdu Schun 60059, hn El: Mo Fng Key Lb of Erth Explorton nd Inforton Technques of Educton Mnstry of hn, hengdu, Schun 60059, hn; ollege of Inforton Engneerng, hengdu Unversty of Technology, hengdu Schun 60059, hn El: Tn Sheng-l ollege of oputer Scence nd Technology, Xuchng Unversty, Xuchng Henn 46000, hn El: Guo X-rong Key Lb of Erth Explorton nd Inforton Technques of Educton Mnstry of hn, hengdu, Schun 60059, hn; ollege of Inforton Engneerng, hengdu Unversty of Technology, hengdu Schun 60059, hn El: Abstrct Reote sensng ge wth chrcterstcs of ultple gry level, ore nfortve, fuzzy boundry, coplex trget structure nd so on, there s no copletely relble odel to gude the reote sensng ge segentton. In response to these ssues, the rtcle presents reote sensng ge sequence segentton ethod bsed on proved FM (fuzzy c-ens) lgorth. The color spce selects the lower relevnce of HSI (hue, sturton, ntensty) nd dopts stndrd covrnce trx the Mhlnobs dstnce forul, whch s ore sutble for the use of reote sensng ge. It cn solve the ntl centers selecton probles of fuzzy -ens clusterng lgorth by the use of EM. By usng the prtton of S coponent, t cn dvde the ge nto hgh S regons nd low S regons. We cn do FM segentton respectvely wth H coponent nd I coponent of these two prts. The segentton results cn be cheved fter the erger. The progr experentl result shows tht ths ethod wll enble FM to converge to globl optl soluton wth less terton, nd hs good stblty nd robustness. It hs good effect on provng the ccurcy of threshold segentton nd effcency for reote sensng ges, whch cn be used for content-bsed reote sensng ge retrevl systes. Index Ters reote sensng ge, fuzzy c-ens, sequence segentton, evolvng clusterng ethod, content-bsed ge retrevl orrespondng uthor, Du Gen-yun, xcdgy@sn.co, I. INTRODUTION Ige segentton s the process nd technology tht to dvde ges nto the regons wth dstnctve fetures nd extrct the trget nterested n. Fetures cn be gry, color, texture nd so on, trget cn correspond to sngle regon, y lso correspond to nuber of regons []. Reote sensng ge segentton ens the process nd technology of the processng, nlyss, extrcted trget. Reote sensng ge wth chrcterstcs of ultple gry level, ore nfortve, fuzzy boundry, coplex trget structure nd so on, whch ke t sk for hgher requreents to reote sensng ge segentton whether n the effcency or the effect, but lck of relble odel to gude t copletely, t block the pplcton of segentton technology n the feld of reote sensng to soe extent [2], [3]. Segentton s n essentl ssue n ge descrpton or clssfcton [4]. Threshold-bsed segentton s bsc ethod of ge segentton -- sply wy nd Hgh-speed del -- but t s not sutble for ge segentton fuzzy boundry regon s ts hrd prtton ethod. Unsupervsed clusterng ethod used regulrly for the fuzzy boundry regon prtton, lke K-Mens, Fuzzy -Mens [5], [6], ISODATA [7], etc [8], [9], [0]. Mny dfferent segentton pproches hve been developed tht cnnot be generlzed under sngle do:0.4304/sw

2 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY schee. The fuzzy -ens lgorth (FM) hs been utlzed n wde vrety of ge processng pplctons such s edcl gng [], [2] nd reote sensng [3], [4]. The lgorth nd vrous odfctons of t wth focus on prctcl pplctons n both ndustry nd scence re dscussed. Its dvntges nclude strghtforwrd pleentton, frly robust behvor, pplcblty to ultchnnel dt, nd the blty to odel uncertnty wthn the dt. FM s ore effectve to the fuzzy boundry regon segent, but the bggest dsdvntge s tht no better wy to deterne the vlue of clusterng nd the ntl cluster centers, essentlly, FM s locl serch optzton lgorth, t wll converge to the locl nu pont nd ths clusterng effect would hve greter pct f the ntl selecton vlue re not properly [5], [6]. In ths pper, new ethod of ge sequence segentton s put forwrd by cobnng the evolvng clusterng ethod wth the proved FM lgorth. The color spce selects the lower relevnce of HSI (hue, sturton, ntensty) nd dopts stndrd covrnce trx the Mhlnobs dstnce forul, whch s ore sutble for the use of reote sensng ge. It cn solve the ntl centers selecton probles of fuzzy - ens clusterng lgorth by the use of EM. By usng the prtton of S coponent, t cn dvde the ge nto hgh S regons nd low S regons. We cn do FM segentton respectvely wth H coponent nd I coponent of these two prts. The segentton results cn be cheved fter the erger. The progr experentl result shows tht ths ethod wll enble FM to converge to globl optl soluton wth less terton, nd hs good stblty nd robustness. It hs good effect on provng the ccurcy of threshold segentton nd effcency for reote sensng ges, whch cn be used for content-bsed reote sensng ge retrevl systes. Ths pper s orgnzed s follows. Secton 2 ntroduces the FM lgorth nd ts probles, nd put forwrd the proved FM lgorth. Secton 3 descrbes our pproch, ncludng color spce, the choce of dstnce esure, nd put forwrd the ethod of reote sensng ge sequence segentton. Results fro the pleented lgorth re shown nd dscussed n Secton 4; whle our conclusons re gven n Secton 5. II. THE IMPROVED FM ALGORITHM A. Fuzzy -ens Algorth Fuzzy -ens [7], [8] s clusterng ethod whch llows pece of dt to belong to two or ore clusters, whch s frequently used n coputer vson, pttern recognton nd ge processng. The FM lgorth obtns segentton results by fuzzy clssfcton [9]. Unlke hrd clssfcton ethods whch group pxel to belong exclusvely to one clss, FM llows pxel to belong to ultple clsses wth vryng degree of ebershps [20]. FM pproch s qute effectve for ge segentton. Severl segentton lgorths bsed on fuzzy set theory nd FM were reported. Fuzzy -ens s clusterng lgorth tht used ebershp degree to deterne ech dt pont belongs to certn cluster. FM dvded the n vectors x (, 2,..., N) nto c fuzzy group, nd coputng the cluster center of ech group, kng vlue functon of non-slrty ndex to cheve the nu. FM lgorth kng ech of the gven dt ponts wth vlues between 0, ebershp to deterne ts degree of belongng to vrous groups through fuzzy prtton. And to sut the ntroducton of fuzzy prtton, the ebershp trx U llowed the eleent vlue between 0,. In ddton to norlzed provdes the ebershp degree su of dt set equvlent to : u,,2,..., N () Then, the vlues functon of FM (obectve functon) s follows: J ( U, v,..., v ) N J u d u rnged between 0, here, v ( ) ( ) 2 (2) s the cluster center of fuzzy group, d x v s the Euclden dstnce between the frst cluster center wth the frst dt pont, [, ) s weghted ndex. The new constructed obectve functon cn be obtned necessry condton so tht (2) to cheve the nu: J ( U, v,..., v,,..., ) c J ( U, v,..., v ) ( u ) N c N N 2 ( u ) ( d ) ( u ) n (3) In (3), the (, 2,..., N ) s Lgrnge ultpler of n of constrned foruls of (). Dervton of ech nput preter, so tht (2) to rech the nu necessry condtons re s follows: nd N N v ( u ) x / ( u ) (4) 2/( ) / k (5) k u d d Fro the bove two necessry condtons, Fuzzy - ens s sple tertve process. In the run-te of btch ode, FM wth the followng steps to deterne the cluster centers v nd the ebershp trx U.

3 30 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY 2009 Step. Usng rndo nuber n vlue between 0, to ntlze the ebershp trx U, to eet constrnts of (). Step2. oputng the cluster centers v (, 2,..., c) by usng (4). Step3. lcultng the vlues functon wth (2). The lgorth to stop f t s less thn deterned threshold or ts prevous vlue of the reltve chnge n functon vlue s less thn certn threshold. Step4. lculton of the new trx U by (5), return step 2. The bove lgorth cn lso ntlze cluster center frstly, nd run the tertve process fter. The lgorth does not gurntee n optl soluton to the convergence, so ts perfornce depends on the ntl cluster centers. Therefore, nother fst lgorth cn be used to deterne the ntl cluster centers, or strt the lgorth wth dfferent ntl clusters center every te, runnng FM repetedly. B. Evolvng lusterng Method Evolvng clusterng ethod s clusterng lgorth of knd of evoluton, on-lne nd boundng by xu dstnce. It ncreses the nuber of cluster or dusts the centers nd the rdus rel-te dynclly s entered sple dt ncresng [2], [22]. In ny one cluster, the xu dstnce between the exple of cluster ponts nd the correspondng xu dstnce re less thn the threshold D threshold, D threshold selecton wll hve drect pct on the clusterng nubers. Exples of clusterng process fro dt stre, the whole process strt clusterng fro n epty set. Soe of the creted clusters to be updted through depend on the locton of the current exple n the nput spce s well s chngng the locton of the cluster centers nd ncresng the rdus of the cluster wth the new exples ppernce, t wll no longer to be updted when ts rdus eet the threshold. RGB (red, green, blue) spce s selecton by spce of ges color -- pre-set the cluster rdus for Dthreshold -- cn pleent the ntl prtton n pxels for rdus of D threshold through the scn of whole ge. The lgorth s: Step. Redng pxel nforton fro ge dt stre n lne, sply select the frst pxel x RGB vlue fro the nput dt stre s clusterng center v of the RGB vlue to crete the frst cluster V, set the cluster rdus r 0. Step2. The lgorth close f ll the pxels re processed n dt stre. Otherwse, coputng dstnce d x v,, 2,..., N between the current nput pxel x, nd N of the cluster center v. Step3. If there s d less thn or equl to the one of r t lest, whch ens tht x belong to the frst cluster V, nely, n( x v ),, 2,..., N d by r d x v, t hs been bounded. Under these crcustnces, nether to crete new cluster nor updte ny exstng clusters. Algorth to return to step 2, otherwse, go nto step 4. Step4. lculte s d r,, 2,..., N, select the clusterng center v tht ws provded wth s, n order to dentfy the cluster V, nd s d r n( s ),, 2,..., N. Step5. x does not belong to ny exstng cluster f s 2 Dthreshold. rete new cluster slr step, then bck to step 2. Step6. Updte V through ove v nd ncrese the vlue of r, f 2 new S Dthreshold. Mke v locted the connecton t x nd v, further, eet new v x r when r s /2 new new bck to step 2.. Optzton of luster enter, the lgorth Ech cluster center fro evoluton lgorth to be optzed by FM, the steps re: Step. Set of fuzzy clusterng coeffcent vlue b nd lgorth ternton threshold, nuber of tertons t, pert the xu nuber of tertons t x, nd c of the cluster centers v ( ) fro evoluton lgorth s the ntl cluster centers for FM. Step2. lculte ebershp functon wth the current cluster center ccordng (5). Step3. Updte ll types of cluster centers wth the current ebershp functon ccordng (4). Step4. Select sutble trx nor, f ( t ) ( t) V V or t tx, then operton stop, otherwse, updte the cluster center vlue t t, return to step 3. When the lgorth convergence, t cn get ech clusterng centers nd the ebershp degrees of vrous clusters sples, copletng the prtton of fuzzy clusterng, nd fnlly through the elnton of fuzzness nto certntes clssfcton, to cheve fnl clusterng segentton. D. Set Prttonng of Ige Pxel Scnnng the dt stre, re-deterne the clssfcton of ech pxel for the c of optzed cluster center. If there s cluster center v, the dstnce

4 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY d r (,..., ), then x belong to cluster V tht t eet the nu dstnce d x v n( x v ). A. olor Model Selecton III. OUR APPROAH Drectly kng use of these coponents wll ffect the results, becuse the RGB (red, green, blue) color spce hs hgh relevnce ong the three coponents. To reduce the pertnence between vrous feture coponents n chrcterstc spce of color use, HSI spce, whch the vson syste bsed on hun percepton of color chrcterstcs of ge processng s dopted n ths pper for the purpose of the convenence specfc pplcton of reote ge segentton. For prctcl ge segentton tsks, the HSI color odel should hve two prncpl dvntges. Frst, the I coponent s de-coupled fro the color nforton n the ge. Second, the H nd S coponents re nttely relted to the wy wth whch hun bengs perceve color [23]. To better tke the dvntges of the color contents of ges, the color ge segentton cn be crred out n HSI spce. The converson fro RGB coponent s to HSI coponents cn be crred out ccordng to the followng foruls. H ( R G) ( R B) rccos 2 ( ) 2 R G ( R B )( G B ) R G or R B, B G, H=2 - H (6) 3 S n( R, G, B) (7) R G B R G B I (8) 3 In the HSI odel spce, every sngle unforty color regon corresponds to reltvely consstent hue H, whch ens tht hue cn be used to conduct segentton n shdow-ndependent color regon. B. Dstnce Mesure Selecton onventonl fuzzy -ens clusterng ethod s bsed on Euclden dstnce,.e. the clusterng ethod of sotropy. However, the ctul sctter of reote sensng ge show tht the dstrbuton of pxel s not subect to sotropc or sphercl dstrbuton. Therefore, the desred results cnnot be often cheved n the pplctons [24], [25]. Frst of ll, the dstrbuton of dfferent ctegores tends to super-ellpsod sctter n the feture spce becuse of the ndeterncy of reote sensng ges nd the exstence of xed pxels, whch re not sutble for the use of dstnce-bsed Euclden dstnce between ponts. Secondly, FM clusterng lgorth cn only be used to, cluster spectru nforton nd does not reflect the dependence between sples n the feld of reote sensng. d x v s equl to the Euclden dstnce nd cn be ppled to the dstrbuton of the se or sphere n the clsscl FM lgorth. Due to the ndeterncy of reote sensng nforton nd the exstence of xed pxels, the dstrbuton of dfferent ctegores tends to super-ellpsod sctter n the feture spce. Thus clusterng results by usng the FM lgorth re not stsfctory. For ths reson, we focus on the chrcterstcs of reote sensng ge nd prove the dstnce lgorth n FM by usng the Mhlnobs dstnce forul, whch s bsed on stndrd covrnce trx nd sutble for the reote sensng ge even ore. Frst of ll, the covrnce trx of three-chnnel nd the stndrd covrnce trx should be clculted. Then the nverse trx of stndrd covrnce trx cn be clculted nd the equton for coputng the dstnce s s follows: 2 ( ) ( ) T d x v ( x v ) (9) In (9), s the nverse trx of the stndrd ( I s unt trx), covrnce trx. When the equton (9) s chnged nto Euclden dstnce. The essence of Mhlnobs dstnce s to ntroduce the weghted trx I n the dstnce clculton of sple X nd the center V. Ths weghted trx s nverse trx of norlzed fuzzy dscrete clss; ntutvely, tht s, the weghted s reltvely less n the bggsh dsperson drecton, wheres n the lesser dsperson drecton weghted s reltvely bgness. In ths wy, we cn cheve the super ellpsod fuzzy clusterng, whch cn ore effectvely detect the dstrbuton of the vrous ctegores n super ellpsod.. Sequence Segentton Sstrtegy Reote sensng ge segentton tkes step-bystep strtegy. For ost proposed lgorths of color segentton usng HSI odel, the three coponents of color ge re tken sultneously n the process. Becuse HSI three coponents re utully ndependent, t s possble to turn ths 3-D serch proble nto three - D serch probles. In the rtcle, the ethod of the sequence segentton of dfferent coponents s dopted nd the flow chrt s shown s n Fg.. () Seprte the ge nto hgh nd low sturton regons by S coponent. (2) Splt hgh sturton regons by usng vlue H, s threshold. In the hgh S regons, the H coponent provdes powerful dscrntng tool for segentton. (3) Splt low sturton regons by usng vlue I, s threshold. In the low S regons, the I coponent hs the coprble blty. Segentton n the bove three steps cn be spltted by usng both dfferent ethods nd the se ethod.

5 32 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY 2009 Hgh S Regons Segentton Usng H Orgnl Reote Sensng Ige RGBà HSI Subdvson Usng S Result obnton Post Processng Segented Reote Sensng Ige Low S Regons Segentton Usng I Fgure. Flow chrt of ge sequentl segentton. II. EXPERIMENTAL RESULTS AND DISUSSION A. Experentl Results wth Iproved FM Algorth Ths pper presents reote sensng ge segentton results for the purpose of verfyng the vldty of the lgorth. Fg.2 s gven segentton results of reote sensng ge bsed on the proved FM lgorth. Fg.2() s rndo sple ge of pxels fro lrge nuber of reote sensng ges, fg.2(b) s the result of the reote sensng ge clusterng segentton bsed on EM nd GA (genetc lgorth), fg.2(c) s the result bsed on EM nd FM. In the experent, ech pxel tke the color of ts cluster center - preters tkes 99.6, the optzed D threshold threshold of n the 2th tertve when the dfference between the old nd the new center vlue s less thn 0.0, to eet the lt of the dfference degree nd wthdrw fro tertve cycle. () orgnl ge (b) result bsed on EM nd GA (c) result bsed on EM nd FM Fgure 2. Segentton results wth proved FM lgorth. Accordng to lrge nuber of experents were found tht usng genetc lgorths cobned wth EM on the se ge segentton to optze the ntl cluster centers fro EM ke use of genetc lgorth, the defnton of evluton functon for the su of ech pxel to t correspondng center dstnce, the end of the condtons of lgorth s optl ndvdul evluton functon for two genertons dfference s less thn 0.0 or the evoluton of the 500 generton, the results showed tht the te the lgorth used ws fr less thn the bove-entoned ethods used n, the dfference s ore obvous when lrger ges nd ore ctegory of ge content. The effect of two ethods s not very dfferent fro rtfcl evluton. As for ge segentton on the se ge wth the se lgorth, t s lso found tht the te spendng on the cobnton of EM nd FM s bsclly stble; whle the te nd the nuber of terton on the cobnton of EM nd GA re not the se. The segentton effect s not very dfferent fro rtfcl udgent when the cluster rdus s pproprtely set. The experent results show tht the segentton effect wth the proved FM lgorth s better when the cluster rdus s pproprtely set. Fg.3 shows the terton nubers curve of two lgorths wth the cluster rdus chnges under the se experentl condtons. The terton nubers for the cobnton of EM GA re tken by coputng the verge for 0 tes wth the se preters. It cn be seen tht the convergence curve s superor to the cobnton of EM nd genetc lgorth, whch ndctes tht lgorth proposed n ths pper s resonble nd effectve. Fgure 3. Algorth convergence curve coprson. B. Experentl Results wth Sequence Segentton In order to verfy the vldty of the bove-entoned ethod, ccordng to the techncl route, HSI spce s selected s the ge color spce nd ore sutble for

6 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY reote sensng ges Mhlnobs dstnce forul s tken s dstnce esure. And then the sequence segentton strtegy s dopted. A Len ge (sze pxels) nd reote sensng ge s selected to crry out the progr experent. The experent results re shown n Fg.4 nd Fg.5. The reote sensng ge dt dopt SPOT5 ult-spectrl ges wth sptl resoluton of 0 nd study re sze of pxels. The proposed lgorth the bove-entoned proved FM lgorth s dopted on the Len ge sequence segentton experents. Fg.4() s the hue dgr for Len ge. Fg.4(b) s the sturton dgr. Fg.4(c) s the ntensty dgr, Fg.4(d) s the thresholdng segentton of hgh S regons n Len ge wth vlue H tht s used the proved FM lgorth. Fg.4(e) s the thresholdng segentton of low S regons n Len ge wth vlue I tht s used the proved FM lgorth. Fg.4(f) s the segentton results fter the erger. () hue dgr (b) sturton dgr (c) ntensty dgr (d) segentton of hgh S regons wth vlue H (e) segentton of low S regons wth vlue I Fgure 4. Sequence segentton results on Len ge. (f) ergng results () hue dgr (b) sturton dgr (c) ntensty dgr (d) segentton of hgh S regons wth vlue H (e) segentton of low S regons wth vlue I Fgure 5. Sequence segentton results on reote sensng ge. (f) ergng results

7 34 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY 2009 The proposed lgorth the bove-entoned proved FM lgorth s dopted on the reote sensng ge sequence segentton experents. Fg.5() s the hue dgr for reote sensng ge. Fg.5(b) s the sturton dgr. Fg.5(c) s the ntensty dgr. Fg.5(d) s the thresholdng segentton of hgh S regons n reote sensng ge wth vlue H tht s used the proved FM lgorth. Fg.5(e) s the thresholdng segentton of low S regons n reote sensng ge wth vlue I tht s used the proved FM lgorth. Fg.5(f) s the segentton results fter the erger. Judgng fro the results, usng the proposed lgorth nd strteges for reote sensng ge segentton cn offer less tertons tes to converge to globl optl soluton. At the se te, t hs good stblty nd robustness. Its good effect of segentton cn prove ccurcy nd effcency of reote sensng ge thresholdng segentton. IV. ONLUSIONS Ige segentton s n portnt reserch re ncludng pttern recognton, ge understndng, coputer vson, nd so on, lthough the use of Fuzzy - ens nd ts proveent ethods wth bref descrpton, esy to pleent, good prtton effect n ge segentton, there re stll soe probles, such s wek robustness of dstnce esure, requreents of settng the ntl nuber of clusters n dvnce, wthout consderng locl ge feture, whch lts ts further pplcton. Moreover, the reote sensng ges wth lrge-scle led to the processng effcency of the lgorth hve certn requreents. In response to these ssues, the rtcle presents reote sensng ge sequence segentton ethod, whch s bsed on the cobnng of the evolvng clusterng ethod wth the odfed FM lgorth. The color spce selects the lower relevnce of HSI (hue, sturton, ntensty) nd dopts stndrd covrnce trx the Mhlnobs dstnce forul, whch s ore sutble for the use of reote sensng ge. It cn solve the ntl centers selecton probles of fuzzy - ens clusterng lgorth by the use of EM. The sequence segentton s crred out n ccordnce wth the strtegy nd the results cn be perfored on the new dgtl globl pltfor. The progr experentl result shows tht ths ethod wll enble FM to converge to globl optl soluton wth less terton, nd hs good stblty nd robustness. It hs good effect on provng the ccurcy of threshold segentton nd effcency for reote sensng ges, whch cn be used for contentbsed reote sensng ge retrevl systes. AKNOWLEDGMENTS Ths reserch hs been supported by open fund proects of the Key Lb of Erth Explorton nd Inforton Technques of Mnstry of Educton (No. 2008DTKF02). REFERENES [] Zhng Yun, Ige Engneerng(II), Ige Anlyss(Second Verson), Beng: Tsnghu unversty press, 2005,pp.73. [2] F. Wng, Fuzzy clssfcton of reote sensng ges, n IEEE Trns. Geosc. Reote Sens, 990, vol.28, no.2, pp [3] Zhng Yng, Fu-L hung, Wng Shtong, Robust fuzzy clusterng-bsed ge segentton, n Appled Soft oputng, 2009, vol.9, no., pp [4] L. nque, G. Forest, L. Lobrd, A clusterng fuzzy pproch for ge segentton, n Pttern Recognton, 2004, vol.37, no.9, pp [5] J.. Bezdek, Pttern Recognton wth Fuzzy Obectve Functon Algorths, New York: Plenu Press, 98. [6] Mn-Shen Yng, Hsu-Shen Ts, A Gussn kernelbsed fuzzy c-ens lgorth wth sptl bs correcton, n Pttern Recognton Letters, 2008, vol.29, no.2, pp [7] J.. Bezdek, A convergence theore for the fuzzy ISODATA clusterng lgorths, n IEEE Trns. Pttern Anl. Mch. Intell, 980, vol.2, no., pp.-8. [8] J.. Bezdek, R. Ehrlch, W. Full, FM: the fuzzy c-ens clusterng lgorth, n oputers nd Geoscences, 984, vol.0, no.2-3, pp [9] L M, R.. Stunton, A odfed fuzzy c-ens ge segentton lgorth for use wth uneven llunton ptterns, n Pttern Recognton, 2007, vol.40, no., pp [0] Hfne ADEL, Zvdovque BERTRAND, FM wth sptl nd ult-resoluton constrnts for ge segentton, n Lecture Notes oputer Scence, 2005, vol.3656, no.0, pp [] D. L. Ph, J. L. Prnce, A. P. Dgher,. Xu, An utoted technque for sttstcl chrcterzton of brn tssues n gnetc resonnce gng, n Interntonl ournl of pttern recognton nd rtfcl ntellgence, 997, vol., no.8, pp [2] J.. Bezdek, L. O. Hll, nd L. P. lrke, Revew of MR ge segentton technques usng pttern recognton, n Med. Phys, 993, vol.20, no.4, pp [3] E. Rgnot, R. hellpp, nd P. Dubos, Unsupervsed segentton of polretrc SAR dt usng the covrnce trx, n IEEE Trns. Geosc. Reote Sensng, 992, vol.30, no.4, pp [4] W. husrong, P. Thtsh, nd Y. Rngsnser, Syntetc perture rdr (SAR) ge segentton usng new odfed fuzzy c-ens lgorth, n Proceedngs of Geoscence nd Reote Sensng Syposu, 2000, vol.2, pp [5] Qn Kun, Xu Mn, Reote Sensng Ige Segentton Bsed on loud Model nd FM, n Geo-Inforton Scence, 2008, vol.0, no.3, pp [6] Wng Xngyng, Wng hunhu, An Adptve FM Ige Segentton Algorth Bsed on the Feture Dvergence, n Journl of Ige nd Grphc, 2008, vol.3, no.5, pp [7] Ynhu Guo, H.D. heng, Yngto Zhng, etc, A new neutrosophc pproch to ge denosng, n Proceedngs of the th Jont onference on Inforton Scences, Publshed by Atlnts Press, 2008:-6. [8] H.D. heng, X.H. Jng, Y. Sun, J.L. Wng, olor ge segentton: dvnces nd prospects, n Pttern Recognton, 200, vol.34, no.2, pp [9] Ingunn Berget, BZrn-Helge Mevk, Torod Næs, New odfctons nd pplctons of fuzzy -ens

8 JOURNAL OF SOFTWARE, VOL. 5, NO., JANUARY ethodology, n oputtonl Sttstcs & Dt Anlyss, 2008, vol.52, no.5, pp [20] X.. Yng, W.D. Zho, Y.F. hen, X. Fng, Ige segentton wth fuzzy clusterng lgorth bsed on Ant-Tree, n Sgnl Process, 2008, vol.88, no.0, pp [2] Ksbov, N.K.; Qun SONG, DENFIS: Dync Evolvng Neurl-fuzzy Inference Syste nd Its Applcton for Te-seres Predcton, n IEEE Trnscton on Fuzzy Syste, 2002, vol.0, no.2, pp [22] Zhng Tng, Lu Jncheng, L Shuwng, Modelng Approch of Dync TSK Model Bsed on Evolvng lusterng Method, n oputer Mesureent & ontrol, 2006, vol.4, no.4, pp [23] Zhng Yun, Yo Yurong, He Yun, olor ge segentton bsed on HIS odel, n Hgh technology letters, 998, vol.4, no., pp [24] HASI Bgn, MA Jnwen, LI Qqng, etc, Iproved Fuzzy -en lssfer nd oprson Study of Its lusterng Results of Stellte Reotely Sensed Dt, n oputer Engneerng, 2004, vol.30, no. pp.4-5,9. [25] QIU Le, LI Guohu, DAI Kexue, Se-supervsed Iproved Fuzzy -Mens lusterng to Reote-sensng Ige, n oputer Applcton Reserch, 2006, vol.23, no.6, pp DU Gen-yun ws born n PR hn, n974. He s currently n Assocte Professor n the college of oputer Scence nd Technology t Xuchng Unversty, nd Ph.D. cnddte n the college of Inforton Engneerng t hengdu Unversty of Technology. He receved B.S. degree n oputer Scence nd Technology fro Henn Norl Unversty, PR hn, n 997, nd the M.E. degree n Sgnl nd Inforton Processng fro hengdu Unversty of Technology, PR hn, n Hs reserch nterests nclude reote sensng ge processng, reote sensng nd coputer networks. He s eber of the F nd IASIT. MIAO Fng ws born n PR hn, n958. He s currently Professor nd doctorl dvsor n the ollege of Inforton Engneerng t hengdu Unversty of Technology. He receved PH.D degree fro hengdu Unversty of Technology. Hs reserch nterests nclude reote sensng nd coputer networks. He s senor eber of the F. TIAN Sheng-l ws born n PR hn, n978. He s currently lecturer n the college of oputer Scence nd Technology t Xuchng Unversty, he receved M.E. degree n ppled thetcs fro Henn Unversty, PR hn, n Hs reserch nterests nclude coputtonl thetcs. GUO X-rong ws born n PR hn, n97. She s Ph.D. cnddte n the college of Inforton Engneerng t hengdu Unversty of Technology. She receved M.E. degree n coputer pplcton fro hengdu Unversty of Technology, PR hn, n Her reserch nterests nclude reote sensng ge processng, reote sensng.