文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 自然科学 > 天文/地理 > A Hybrid Object-Oriented Conditional Random Field
侵权投诉

A Hybrid Object-Oriented Conditional Random Field

遥感影像分类英文文献

spectralvariability[24].However,theclassi cationresultisdirectlyaffectedbythesegmentationscale,whichisimportantbutchallengingtochoose[25].Forsingle-scalesegmentation,thedif cultyofchoosingthescaleisthatthereisnopriortodeterminewhatparameterswillproduceagoodresult,andvaryinglandcovertypesoftenhavedifferentoptimalscalesduetothedifferenceinsize[17].

Afurtherinterestingapproachistherandom eldmethodrepresentedbyMarkovrandom elds(MRFs),whichnotonlyincorporatesthespatialcontextualinformationbutavoidsthescalechoiceproblemformodelingthespatialdependenciesofthepixels.MRFwas rstintroducedintoimageprocessingin1984[26]andhasbeenwidelyusedinclassi cationproblemsinrecentyears[3],[27]–[29].Inaddition,anovelframeworkcalledMarkoviansupportvectorclassi er(MSVC),integratingSVMandMRFmodelsinauniqueformulationforspatialcon-textualclassi cation,hasbeenrecentlyproposed[30]andhasperformedwell.However,MRFconsidersthespatialinforma-tiononlyinthelabelimage,notintheoriginalobservedimagedata.AnimprovedmodelforMRFistheconditionalrandom eld(CRF)model,whichdirectlymodelstheclassposteriorprobability,givenobserveddata,ratherthantheirjointdistribu-tion.Thisapproachhastheabilitytoincorporatethespatialin-formationinboththelabelsandobservedimagedata.CRFwas rstproposedbyLaffertyetal.[31]forsolvingthelabelingof1-Dtextsequencesandhasbeensuccessfullyappliedinimagesegmentation[32],stereovision[33],andactivityanalysis[34]afterbeing rstintroducedintoimageanalysisbyKumarandHebert[35],[36].Theimageclassi cationproblemistypicallyformulatedaspairwiseCRF,modelingthespatialdependenciesinthelocalneighborhoods(wheretheneighborhoodistypicallyde nedasa4-or8-neighborhoodinthepixelgrid),whichhasbeensuccessfullyappliedtohyperspectralimagery[37]–[39].However,theresultsofthismodelalwayshaveanoversmoothappearance[39],[40].Althoughsometimestheoversmoothingphenomenonisnotobviousincertainimages,itisimportanttoalleviatetheeffectforHSRimageryduetothepresenceofsmallimportantstructures.Inordertocopewiththisproblem,high-orderpotentials[40],[41],modelingthemorecomplexstatisticsoftheimage,havebeenusedandhaveachievedgoodperformancesinexperiments.However,itisdif culttoachieveanef cientinferencewiththesehigh-orderpotentials,andtheyalwaysneedtoimportlocalorglobalinformation,whichincreasesthecomplexityofthemodel.

Inthispaper,ahybridobject-orientedCRFclassi cationframeworkforHSRimagery(CRF+OO)isproposedtoutilizeboththepairwiseCRFmodelandtheobject-orientedclassi -cationmethod,toincorporatethespatialcontextualinformationforHSRimageryclassi cation.TheCRF+OOclassi cationframeworkisdescribedinthefollowing.

1)ACRFclassi cationframeworkcontainingtwoCRFmodelsisdesigned.ThetwoCRFmodelshavethesamepairwisepotentialtoconsidertheneighborhoodinter-actionsbutdifferentunarypotentials,whicharecalledthelogunarypotentialandthequasi-gammaunarypo-tential,respectively;thus,thecorrespondingmodelsarerepresentedasCRF-LOGandCRF-QG,respectively.The

unarypotentialsuseprobabilisticpixelwiseclassi cationtomodeltherelationshipbetweentheobserveddataandthecorrespondinglabel.However,thequasi-gammaunarypotentialinCRF-QGisdesignedtogivealargerweighttothespectralinformation,toimposearestric-tionontheless-possiblepixellabelsandtofavorthemostlikelylabelsofthelandcovertypes.Therefore,CRF-LOGandCRF-QGalwayshavedifferentdegreesofsmoothclassi cationperformance,andtheyprovidecomplementaryinformationaboutthelandcovertypesincetheyalsopossessthesamepairwisepotential,whichallowsthemtohavetheabilitytoincorporatespatialinformation.

2)AlthoughthepotentialshavebeenmodeledintheCRFclassi cationframework,theinferencesearchingforthebestsolutioncorrespondingtotheoptimalpixellabelingisanNP-hardproblem[42]forthemultivaluedvariablesinHSRimageryclassi cation.To ndtheoptimallabel-ing,differentapproaches,e.g.,iteratedconditionalmodes(ICMs),loopybeliefpropagation(LBP),andgraphcuts,havebeenproposed.However,ICMeasilygetsstuckatpoorlocalminima;thus,itisextremelysensitivetotheinitialestimate,andLBPisnotguaranteedtoconverge.Inthispaper,thegraph-cut-basedα-expansionalgorithm[43]isappliedtotheHSRimageryclassi cationsinceithasabetterperformance(ef ciencyandaccuracy)incomputervision[44].

3)Toobtainamorerobustandexcellentclassi cationresult,aclassi cationfusionschemeisproposedtomakefulluseoftheformerclassi cationinformation.SincetheresultsofthetwoCRFmodelsincorporatethespatialcon-textualinformation,theCRF-LOGandCRF-QGmapshavetheabilitytoobtainhomogeneousregions.There-fore,theyareselectedtoobtainthesegmentationresultbytheconnected-componentlabelingalgorithmsothatwecanmakeafusionoftheclassi cationinformationattheobjectlevel.Inaddition,theclassi cationmapcanprovidepriorclassinformationtoaddressthescalechoiceproblembecausetheclassi cationresultre ectsthedistributioninformationofthematerialsofinterest,sothatthescaleofthesegmentationmeetstherequirementofallthelandcoverclasses.The nalclassi cationresultisachievedbyintegratingtheobject-orientedclassi ca-tions,usingamajorityvotingstrategywitharegionsizeconstraintattheobjectlevel.Theef ciencyoftheproposedCRF+OOclassi cationframeworkiscon rmedbyperformingexperimentsonthreedatasets,whichconsistsoftwomultispectralHSRimages(QuickBirdandIKONOS)andahyperspectralHSRimage(HYDICE).Comparedwithotherstate-of-the-artalgorithms,theproposedalgorithmhasaremarkablequantitativeandqual-itativeperformance.

Theremainderofthispaperisorganizedasfollows.ThegeneralMRFandCRFmodelsarebrie ypresentedinSectionII.InSectionIII,wegiveadetaileddescriptionoftheCRF+OOclassi cationframeworkforHSRimagery.SectionIVprovidestheexperimentalresults,andthesensitivity

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email