資源描述：
《Normalized Cuts and Image Segmentation》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、888IEEETRANSACTIONSONPATTERNANALYSISANDMACHINEINTELLIGENCE,VOL.22,NO.8,AUGUST2000NormalizedCutsandImageSegmentationJianboShiandJitendraMalik,Member,IEEEAbstractDWeproposeanovelapproachforsolvingtheperceptualgroupingprobleminvision.Ratherthanfocusingonlocalfea
2、turesandtheirconsistenciesintheimagedata,ourapproachaimsatextractingtheglobalimpressionofanimage.Wetreatimagesegmentationasagraphpartitioningproblemandproposeanovelglobalcriterion,thenormalizedcut,forsegmentingthegraph.Thenormalizedcutcriterionmeasuresboththe
3、totaldissimilaritybetweenthedifferentgroupsaswellasthetotalsimilaritywithinthegroups.Weshowthatanefficientcomputationaltechniquebasedonageneralizedeigenvalueproblemcanbeusedtooptimizethiscriterion.Wehaveappliedthisapproachtosegmentingstaticimages,aswellasmoti
4、onsequences,andfoundtheresultstobeveryencouraging.IndexTermsDGrouping,imagesegmentation,graphpartitioning.?1INTRODUCTIONNEARLY75yearsago,Wertheimer[24]pointedoutthetobedonefromthebigpicturedownward,ratherlikeaimportanceofperceptualgroupingandorganizationpaint
5、erfirstmarkingoutthemajorareasandthenfillingininvisionandlistedseveralkeyfactors,suchassimilarity,thedetails.proximity,andgoodcontinuation,whichleadtovisualPriorliteratureontherelatedproblemsofclustering,grouping.However,eventothisday,manyofthegroupingandimag
6、esegmentationishuge.Theclusteringcomputationalissuesofperceptualgroupinghavere-community[12]hasofferedusagglomerativeanddivisivemainedunresolved.Inthispaper,wepresentageneralalgorithms;inimagesegmentation,wehaveregion-basedframeworkforthisproblem,focusingspec
7、ificallyonthemergeandsplitalgorithms.Thehierarchicaldivisivecaseofimagesegmentation.approachthatweadvocateproducesatree,thedendrogram.SincetherearemanypossiblepartitionsofthedomainIWhilemostoftheseideasgobacktothe1970s(andearlier),ofanimageintosubsets,howdowe
8、pickthearightoone?the1980sbroughtintheuseofMarkovRandomFields[10]Therearetwoaspectstobeconsideredhere.Thefirstisthatandvariationalformulations[17],[2],[14].TheMRFandtheremaynotbeasingleco
