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

ZHONGetal.:OBJECT-ORIENTEDCRFCLASSIFICATIONFRAMEWORKFORHSRREMOTESENSINGIMAGERY7025

analysisisdiscussedinSectionV.Inthe nalsection,theconclusionisgiven.

II.MARKOVANDCONDITIONALRANDOMFIELDMODELSInprobabilisticschemes,undirectedgraphicalmodels,alsoreferredtoasrandom elds,havebeenusedtoincorporatespatialcontextualinformationincomputervision.AsdescribedinSectionI,thisrandom eldmethodofmodelingtheneigh-borhoodinteractionsisveryimportantinHSRremotesens-ingimageclassi cationandcanimprovetheperformanceofpixelwiseclassi cation.MRFisthemostpopularundirectedgraphicalmodel,incorporatingthelocaldependenciesbetweenrandomvariablesusingprobabilisticframeworks[27],[36].IntheMRFclassi cationframework,theposteriorprobabilityofthelabels,giventheimagedata,isdescribedinthefollowing,whichisbasedonBayes’rule:

P(x|y)∝P(x,y)=P(x)P(y|x)

(1)

wherexrepresentsthecorrespondinglabelsofthewholeimage,andyaretheobserveddatafromaninputimage.TheMRFframeworkmodelsthejointprobabilityoftheobserveddataandthecorrespondinglabelsP(x,y).InMRF,P(x)isthepriordistributionofthelabelsx,whichisformulatedasaGibbsdistributiontoconsiderthespatialinformation,i.e.,

P(x)=

1

Z

exp ψc(xc)

(2)

c∈C

whereZ=

xexp{ c∈Cψc(xc)},whichisusuallynamedthepartitionfunction,isanormalizationfactor,andthetermψc(xc),whichiscalledthepotentialfunction,islocallyde nedinthecliquec,whichisasubsetofvariablesxc x.Cisthesetofallthecliques.Forcomputationaltractability,theobserveddataareassumedtobeconditionallyindependent,andthelikelihoodP(y|x)=

P(y|x)canhaveafactorizedform,i.e.,P(y|x).However,itshouldbenotedthatthespatialcontextualinteractionmodeledbythetermP(x)isrestrictedtothelabeling eldanddoesnotdependontheobserveddata.

ComparedwithgenerativeMRFexpendingeffortstomodelthejointdistributionP(x,y),CRFdirectlymodelstheposte-riorprobabilityofthelabels,giventheimagedataP(x|y)[31],[36]withthefollowingform,whichisthatwewanttoestimateintheclassi cationtask:

P(x|y)=

1

Z(y)

exp ψc(xc,y)

(3)

c∈C

whereZ(y)=

exp{ function,andthetermxψc∈Cψc(xc,y)}isthepartitionc(xc,y)denotesthepotentialfunction.Intheclassi cationproblem,themostcommonCRFmodelisknownaspairwiseCRF,whichmodelsthespatialdependenciesofpairsofrandomvariablesinthelocalneighborhoods.Intheory,thepotentialfunctionscanincludeunarypotentials,

pairwisepotentials,andevenhigh-orderpotentials,basedonthedifferenttypesofcliquesintheobserveddataandtheircorrespondinglabels.CRFcanincorporatemorewide-rangingcontextualinformation,asde nedbythehigh-orderneigh-borhoodsystemandthecliquesforthehigh-orderpotentials.However,optimizationmethodsforthegeneralcaseareinfea-sible.Therefore,inthispaper,weusepairwiseCRFwithan8-neighborhood.

ItcanbeseenfromtheformulationsofMRFin(1)and(2)andCRFin(3)thatboththeMRFandCRFmodelshavetheabilitytoconsiderthespatialcontextualinformationinthelabel elds,buttheCRFmodelalsohastheabilitytopermitinteractionsintheobserveddata.TheCRFmodeldirectlymodelstheposteriordistributionasaGibbs eld,whichallowsthemodeltoincorporatethecontextualinformationinamore exibleway.Theclassi cationproblemistypicallyformulatedaspairwiseCRF,whichhasbeensuccessfullyappliedinvariousimageanalysis elds[32]–[34],[39],[40].However,theresultsofthismodelwillalwayshaveanoversmoothappearance,tosomedegree,intheclassi cationapplication.Therefore,toutilizethepairwiseCRFmodelandalleviatetheoversmooth-ing,ahybridobject-orientedCRFclassi cationframework(CRF+OO)isusedforthespatialcontextualclassi cationofHSRimageryinthispaper.

III.HYBRIDOBJECT-ORIENTEDCRFCLASSIFICATIONFRAMEWORK

Inthispaper,inordertomakefulluseofthespatialcontextualinformationfortheclassi cationofHSRremotesensingimagery,ahybridobject-orientedCRFclassi cationframework(CRF+OO)isproposedbyintegratingCRFandanobject-orientedclassi cationmethod,whichbothincorporateinformationonthespatialcontextofeachpixel.TheproposedCRF+OOforHSRimagerycanbedescribedinthreemainsteps,asshowninFig.1.

Intheprobabilisticpixelwiseclassi cationstep,fortheinputHSRimageryandthetrainingsamples,aprobabilisticpixelwiseclassi cationis rstperformedtoobtaintheinitialclassi cationresultandthecorrespondingprobabilitymap.Inthispaper,theprobabilisticSVMclassi erisappliedforthispurpose.

TheCRFclassi cationframeworkstep,asthesecondstep,isperformedbycombiningthespectralandspatialinformation,thegoalofwhichistoobtaintheclassi cationresultofCRFbycarryingoutthegraph-cut-basedα-expansionalgorithm.Thissteptakestheprobabilitymapfromthe rststepandtheoriginalHSRimageryastheinputandconsistsoftwoCRFmodels,whichhavethesamepairwisepotentialbutdifferentunarypotentials,whicharecalledthelogunarypotentialandthequasi-gammaunarypotential,respectively.ThecorrespondingmodelsarerepresentedasCRF-LOGandCRF-QG,respec-tively,intheclassi cationframework,andtheyexporttwocor-respondingclassi cationmaps.SinceCRF-LOGandCRF-QGbothhavetheabilitytoincorporatespatialinformationbutpossessdifferentunarypotentials,theyalwayshavedifferentdegreesofsmoothclassi cationperformanceandcanprovidecomplementaryinformationaboutthelandcovertypes.

上一页第3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