9-1-图像分割-概述

图像分割图像识别与人工智能研究所,多谱信息处理国家重点实验室陶文兵华中科技大学图像识别与人工智能研究所多谱信息处理技术国家重点实验室分割的目的和意义图像分割是计算机视觉研究中的基础问题和难点之一图像分割就是把图像分成各具特性的区域并提取出感兴趣目标图像分割的难点和挑战性对一般图像中的大量视觉模型进行建模的复杂性图像理解本身的内在模糊性当没有一个明确的任务来指导注意机制2图像理解图像分析图像处理目标识别特征提取图像分割原始图像高低语义抽象程度小大操作对象数据语义符号目标描述原始像素图像工程的三层模型imagesegmentation…Goal:BreakuptheimageintomeaningfulorperceptuallysimilarregionsTypesofsegmentationsOversegmentationUndersegmentationMultipleSegmentationsSegmentationasaresultRotheretal.2004Segmentationforefficiency[FelzenszwalbandHuttenlocher2004][Hoiemetal.2005,Mori2005][ShiandMalik2001]SegmentsasprimitivesforrecognitionJ.TigheandS.Lazebnik,submittedtoECCV2010MajorprocessesforsegmentationBottom-up:grouptokenswithsimilarfeaturesTop-down:grouptokensthatlikelybelongtothesameobject[LevinandWeiss2006]Bottom-upsegmentation•Grouptogethersimilar-lookingpixelsforefficiencyoffurtherprocessing“Bottom-up”processUnsupervisedX.RenandJ.Malik.Learningaclassificationmodelforsegmentation.ICCV2003.“superpixels”Thegoalsofsegmentation•Separateimageintocoherent“objects”“Bottom-up”or“top-down”process?Supervisedorunsupervised?Berkeleysegmentationdatabase:E.BorensteinandS.Ullman,“Class-specific,top-downsegmentation,”ECCV2002A.LevinandY.Weiss,“LearningtoCombineBottom-UpandTop-DownSegmentation,”ECCV2006.Top-downsegmentationE.BorensteinandS.Ullman,“Class-specific,top-downsegmentation,”ECCV2002A.LevinandY.Weiss,“LearningtoCombineBottom-UpandTop-DownSegmentation,”ECCV2006.NormalizedcutsTop-downsegmentation图像分割方法的发展现状目前图像分割方法主要有三个比较重要的方向：基于统计理论的图像分割方法MeanShift，DDMCMC…基于变分模型的图像分割方法SnakeModel，GAC，M-SModel，C-VModel…基于图论的图像分割方法GraphCuts，NormalizeCuts…14ThreebasictheoryinImageSegmentationStatisticsVariationalGraphTwobasicModelinImageSegmentationStatisticsformulationVariationalModelEnergyModelsOptimizationMethodBayesianformulationBayesianformulation(GemanandGeman,1984)Geman,S.andD.Geman:1984,“Stochasticrelaxation,Gibbsdistributions,andtheBayesianrestorationofimages”.IEEETransactionsonPatternAnalysisandMachineIntelligence6,721–741.(13641),(,)()()()()(,)datasmoothpppqpqppqELELELDLVLLPNDpisadatapenaltyfunction,Vp,qisaninteractivefunctionDatapenaltiesindicateindividuallabel-preferencesofpixelsbasedonobservedintensitiesandprespecifiedlikelihoodfunction.Interactionfunctionencouragespatialcoherencebypenalizingdiscontinuitiesbetweenneighboringpixels.MAP-MRF:Maximumaposteriori-MarkovrandomfieldSnakemodel(Kassetal.,1988)Kass,M.,A.Witkin,andD.Terzopoulos:1988,“Snakes:Activecontourmodels”.InternationalJournalofComputerVision,vol.1,pp.321–331(13622)SnakemodelInternalenergyThefirsttwotermscontrolthesmoothnessofthecontourstobedetected.ExternalenergyThethirdtermisresponsibleforattractingthecontourtowardstheobjectintheimage.Geodesicactivecontoursmodel(GAC)(Casellesetal.,1997)V.Caselles,R.Kimmel,andG.Sapiro,“Geodesicactivecontours,”Int.J.Comput.Vis.,vol.22,pp.61–79,1997.(3709)1/(1)gI()0()(())LCbECgCsdsRemarks:Thefunctiongisanedgeindicatorfunctionthatvanishesatobject.Theshorterthecurve,thelargerthegradientofthecurve,thelesstheenergy.Edge-basedActiveContourGeodesicactivecontoursmodelMumfordandShahfunctional(MumfordandShah,1989)Mumford,D.andJ.Shah:1989,‘Optimalapproximationsbypiecewisesmoothfunctionsandassociatedvariationalproblems’.CommunicationsonPureandAppliedMathematics42,577–685.(3122)Mumford-Shahfunctional220\(,)()MSCFuCuudxudxCRemarks:TheminimizationofMumford-ShahfunctionalresultsinanoptimalcontourCthatsegmentsthegivenimageu0intoseveralregions.Imageuisanoptimalpiecewisesmoothapproximationofthegivenimageu0ImageuissmoothwithineachoftheconnectedcomponentsintheimagedomainseparatedbythecontourC.CVModel-piecewiseconstantMSmodel(ChanandVese,2001)T.ChanandL.Vese,“Activecontourswithoutedges,”IEEETrans.ImageProcess.,vol.10,no.2,pp.266–277,Feb.2001.Citedtimes:4514Region-basedActiveContour122212101202(,,)((,))((,))CVFccCuxycuxycdsRemarks:Assumethatuisapiecewiseconstantfunction.Forsuchcase,thesecondtermdisappearsfromtheMSfunctions.Twophaseproblem,c1istheaverageofregion1,c2istheaverageofregion2.Chan-Vese(CV)modelOptimizationMethod1、模拟退火(simulatedannealing)2、水平集算法(LevelSet)3、图割算法(Graphcuts)4、期望最大化算法(Expectation-Maximization—EM)5、置信传播(Beliefpropagation)4、对偶算法(primaldual)