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-ORIENTEDCRFCLASSIFICATIONFRAMEWORKFORHSRREMOTESENSINGIMAGERY

A Hybrid Object-Oriented Conditional Random Field

7027

Fig.2.Relationshipbetweenlikelihoodandenergyvalue.(a)Logunarypotential.(b)Quasi-gammaunarypotential.

energy,whichhasbeenwidelyusedforremotesensingimageclassi cation,canbewrittenasthesumoftheunaryandpairwisepotentials,i.e.,

ψi(xi)+λψij(xi,xj)(6)E(x)=

i∈V

i∈V,j∈Ni

whereψi(xi)istheunarypotentialterm,andψijisthepairwise

potentialterm,whichiscomputedoverthelocalneighborhoodNiofpixeli.Thenonnegativeconstantλtradesoffthestrengthofthepairwisepotentialagainsttheunarypotential.

Aftertheestablishmentofthebasicmodel(6),aremainingproblemishowtoformulatethetwokindsofpotentials(i.e.,unaryandpairwisepotentials)andobtainthe nallabelsbyinference.

1)UnaryPotentials:Unarypotentialsdescribethecostofasinglepixeltakingaparticularlabel,whichdependsonthelocalappearancefeaturesderivedfromtheimage.Typically,theunarypotentialψi(xi)isde nedas[3],[32]

ψi(xi)= ln(P(xi=lk))

(7)

whereP(xi=lk)istheprobabilityofpixelxitakingthelabellk,whichcanbegivenbyvariousdiscriminativeclassi- ers.Inourwork,theseprobabilityestimatesarefrom“one-versus-one”SVMoutputs,asdescribedpreviously.Inordertoconvenientlydescribethedetailsofthealgorithm,wecalltheunarypotentialthe“logunarypotential”becauseofitslogfunctionform.

Theunarypotential’stermisrelatedtothepixelwisein-formation,andtheminimizationofonlythisterm’senergycontributionbyitselfwouldbeequivalenttoanoncontextualBayesianclassi cationoftheimage.Therefore,itshouldbelowwhenthecorrespondingpixeliscorrectlyclassi edandhighwhenitismisclassi ed.Thelogunarypotentialisfor-mulatedasthenegativeloglikelihoodofthepixeltakingtherelatedclasslabels.However,ifasmallobjectcontainingafewpixelswithhighlikelihoodispresentintheimage,thesepixelstendtobeassignedtheclasslabelofthesurroundingobjectsbecausethelogunarypotentialdoesnothaveenoughabilitytopreventthesmoothingofthepairwisepotentials.Inordertosuppresstheoversmoothingoftheclassi cationresult,thequasi-gammaunarypotentialisde ned

ψi(xi)=γ1/P(xi=lk) γ.

(8)

Fig.2(a)and(b)showstherelationshipbetweenthelikeli-hoodandtheenergyvalueofthelogandquasi-gammaunarypotentialswhenγ=2,respectively.Aswiththelogunarypotential,thequasi-gammaunarypotentialisalsomodeledasadecreasingfunctionoflikelihood.However,comparedwiththelogunarypotential,whentheprobabilitydensityestimateofthepixeltakingthecorrespondingclasslabelishigh,thepe-nalizationofthequasi-gammaunarypotentialisstilllow.Thismeansthatthelandcoverclasslabelofthepixelismaintainedinminimizingtheenergyfunctionwhenthecon denceofthepixeltakingtheclasslabelishigh.Inaddition,thepenalizationismuchgreaterwhenthelikelihoodisverylow,andevenwhenthelikelihoodis0.1,thispenalizationcanreach1024,whichimpliesthatthelandcoverclasslabelofthepixelisnotexpectedtobekeptwhenthecon denceisless.

2)PairwisePotentials:Pairwisepotentialsmodelasmooth-nesspriorthatencodesthefactthat,inrealimages,neighboringpixelsinhomogeneousimageregionsusuallytakethesamelabel.AsistypicalforpixellabelingwithCRF,thecontrast-sensitivesmoothnesspriorψij(xi,xj)takestheformof[32],[47],[48]

0,ifxi=xj

(9)ψij(xi,xj)=g(i,j)

,otherwisewhere

g(i,j)=1+θvexp θw xi xj 2.

Inthispaper,theedgefeaturefunctiong(i,j)isdesignedtomeasurethedifferenceinappearancebetweentheneighboringpixels.Theconstantincludeding(i,j)isabiastermtobelearnedforremovingsmallisolatedregions;thepairof(i,j)representsthespatiallocationofneighboringpixels;parameterθvisaconstantdeterminingthedegreeofsmoothness,andthequantityθwisthemeansquaredifferencebetweenthespectralvectorsoveralltheadjacentpixelsintheimage.

Thepairwisepotentialsincorporatethespatialcontextualinformationofeachpixel,whichisexpressedintermsofitsneighborhood,tobetakenintoaccountintheclassi cation.WhenminimizingtheenergyfunctionE(x),thecontrast-sensitivesmoothnesspriortermpenalizesthespatialinconsis-tenciesamongneighboringpixelswithdifferentclasslabels,whilefavoring,intheoutputclassi cationmap,thesameland

上一页第5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