資源描述：
《markov random fields and applications》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、MarkovRandomFieldsandApplicationsClicktoeditMastersubtitlestyleSoumyaGhosh09年3月25日Outline?MRFbasics–BayesianImageAnalysis–MarkovRandomFieldstheory–Gibbs–MarkovEquivalence–Inference–Learning?Application–ImageSegmentation09年3月25日BayesianImageAnalysisNoiseTransmissionOriginalImageDegraded(observ
2、ed)Image09年3月25日BayesianImageAnalysis?LetXbetheobservedimage={x1,x2…xmn}?LetYbethetrueimage={y1,y2…ymn}?Goal:findY=y*={y1*,y2*…}suchthatP(Y=y*
3、X)ismaximum.?LabelingproblemwithasearchspaceofLmn–Listhesetoflabels.–m*nobservations.3/25/09FirstGuess?Assumethelabelsyiareindependent.P(Y
4、X)=∏mni=1P(
5、yi
6、xi)MaximizingP(Y
7、X)boilsdowntosimplymaximizingtheindividualP(yi)s.3/25/09UnfortunatelyObservedSVMMRFImage3/25/09MarkovRandomFields?Introducedinthe1960s,aprincipledapproachforincorporatingcontextinformation.?Incorporatingdomainknowledge.?WorkswithintheBayesianframework.?Widelyworkedoninthe7
8、0sdisappearedoverthe80sandfinallymadeabigcomebackinthelate90s.3/25/09MarkovRandomField?RandomField:LetbeafamilyF={F,F,...,F}12MofrandomvariablesdefinedonthesetS,inwhicheachrandomvariabletakesavalueinalabelsetL.ThefamilyFiscalledarandomfield.Ffii?MarkovRandomField:FissaidtobeaMarkovrandomfield
9、?onSwithrespecttoaneighborhoodsystemNifandonlyifthefollowingtwoconditionsaresatisfied:Possitivity:P(f)>0,"f?FMarkovianity:P(f
10、S-{i})=P(f
11、f)iiNi3/25/09ImageAnalysis?Wecouldthusrepresentboththeobservedimage(X)andthetrueimage(Y)asMarkovrandomfields.X–observedimageY–trueimage?AndinvoketheBayesian
12、frameworktofindP(Y
13、X)3/25/09Details?P(Y
14、X)proportionaltoP(X
15、Y)P(Y)–P(X
16、Y)isthedatamodel.–P(Y)modelsthelabelinteraction.?NextweneedtocomputethepriorP(Y=y)andthelikelihoodP(X
17、Y).3/25/09Cliques?AcliqueisdefinedasasubsetofsitesinF,whereeverypairofsitesareneighborsofeachother.Thecollectionsofsingl
18、e-site,double-site,andtriple-sitecliquesaredenotedbyC1,C2,C3and…Cn3/25/09Markov–GibbsEquivalence?GibbsDistribution:P(f)=(1/Z)e(-1/T)U(f)whereU(f)=∑cinCVc(f)=∑iinC1V1(fi)+∑iinC2V2(fi,fj)+…Z–NormalizingFactoroverthespaceofallconfigurations.?Ham
