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

7026IEEETRANSACTIONSONGEOSCIENCEANDREMOTESENSING,VOL.52,NO.11,NOVEMBER

A Hybrid Object-Oriented Conditional Random Field

2014

Fig.1.Flowchartofthehybridobject-orientedCRFclassi cationframework.

Intheclassi cationfusionstep,thesegmentationis rstper-formedbytheconnected-componentlabelingalgorithm,usingtheCRF-LOGandCRF-QGclassi cationresultsobtainedinthesecondstep.Object-orientedclassi cationmapscanbethenobtainedbyamajorityvotingstrategy,usingthesegmentationmapandtheclassi cationmap.The nalclassi cationresultisachievedbyintegratingtheformerobject-orientedclassi ca-tionsattheobjectlevel.

twoclasses.Finally,intheclassi cation,amajorityvotingstrategyisused.Moredetailscanbefoundin[45].B.CRFClassi cationFramework

Consideranorderedsetofvariablesy={y1,y2,...,yN},whichdenotesthesetofspectralvectorsofanHSRimage,whereNisthetotalnumberofpixelsoftheimage,andyiisthespectralvectoroftheimagepixeli∈V={1,2,...,N}withlengthd(thenumberofspectralbands).Theclassi cationimageisdenotedbyx={x1,x2,...,xn},wherexiinxtakesthevalueofthelabelsetL={1,2,...,K}.Kisthenumberofclasses.

UnderBayes’framework,theimageclassi cationcancor-respondto ndingthemaximumaposteriori(MAP)estimateofthelabelimage,suchthatxMAP=argmaxxP(x|y).IfxconditionedonymeetstheMarkovproperty,CRFdirectlymodelstheposteriorprobabilityasaGibbsdistributionin(3).ThecorrespondingGibbsenergy[40],[41]isde nedas

ψc(xc,y).(4)E(x)= logP(x|y) logZ(y)=

c∈C

A.ProbabilisticPixelwiseClassi cation

The rststepoftheproposedframeworkconsistsofper-formingaprobabilisticpixelwiseclassi cationoftheHSR

imagery.Variousclassi ersthathaveaprobabilisticoutputcanbeused.However,theSVMclassi erissuggestedduetoitshighcapacityandgreateradaptability.Inotherwords,whetherinthecaseofhighdimensionsandasmalltrainingsetornot,SVMalwaysshowsagoodperformance.MoredetailsabouttheSVMclassi ercanbefoundin[45].Inthisstep,theclassi cationmapandthecorrespondingprobabilitymapwiththeposteriorprobabilitiesthatthesamplesbelongingtotheclassesareexpectedtoobtain.However,theoutputofgeneralSVM,viathedecisionfunction,isonlyaclasslabelwithoutprobabilityinformation.Inordertoobtaintheprobabilityes-timates,Platt’sformulationisusedtomaptheoutputintotheprobability[45],[46].Then,theremainingproblemishowtoexpandthealgorithmfrombinaryclassi cationtomulticlassclassi cation.The“one-against-one”approachissuggestedin[45].Assumingthatkisthenumberofclasses,weconstructk(k 1)/2classi erssothateachclassi ertrainsdatafrom

Therefore,theMAPlabelingxMAPoftherandom eldisgivenby

xMAP=argmaxP(x|y)=argminE(x).

x

x

(5)

Equation(5)showsthatthemaximizationoftheposteriorprobabilityP(x|y)isequivalenttotheminimizationoftheenergyfunctionE(x).AcommonlyusedexampleofaCRF

上一页第4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