Learning_a_classification_model_for_segmentation

《Learning_a_classification_model_for_segmentation》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、LearningaClassificationModelforSegmentationXiaofengRenandJitendraMalikComputerScienceDivisionUniversityofCaliforniaatBerkeley,Berkeley,CA94720fxren,malikg@cs.berkeley.eduAbstractWeproposeatwo-classclassificationmodelforgroup-ing.Humansegmentednaturalimage

2、sareusedaspositiveexamples.Negativeexamplesofgroupingareconstructed(a)(b)(c)byrandomlymatchinghumansegmentationsandimages.Inapreprocessingstageanimageisoversegmentedintosu-Figure1.Weformulatesegmentationasclassificationperpixels.Wedefineavarietyoffeatures

3、derivedfromthebetweengoodsegmentations(b)andbadsegmenta-classicalGestaltcues,includingcontour,texture,bright-tions(c).WeuseGestaltgroupingcuesasfeaturesnessandgoodcontinuation.Information-theoreticanaly-andtrainaclassifier.Humansegmentedimagesaresisisapp

4、liedtoevaluatethepowerofthesegroupingcues.usedasexamplesofgoodsegmentations.Badseg-Wetrainalinearclassifiertocombinethesefeatures.Tomentationsareconstructedbyrandomlymatchingademonstratethepoweroftheclassificationmodel,asimplehumansegmentationtoadifferent

5、image.algorithmisusedtorandomlysearchforgoodsegmenta-tions.Resultsareshownonawiderangeofimages.boundaries.Theprincipleofsimilarityistwofold:1.Introduction1.intra-regionsimilarity:theelementsinaregionaresimilar.Thisincludessimilarbrightness,similartex-Pe

6、rceptualgroupingcanbeformulatedasanopti-ture,andlowcontourenergyinsidetheregion;mizationprobleminanumberofdifferentframeworks,2.inter-region(dis)similarity:theelementsindifferentsuchasgraphpartitioning[22,16,8,5]orvariationalregionsaredissimilar.Thisint

7、urnincludesdissimilarapproaches[13].Theobjectivefunctionbeingoptimizedisbrightness,dissimilartexture,andhighcontourenergytypicallydrivenbythedesigner'sintuitionorcomputationalonregionboundaries.convenience.Thethemeofthispaperistoderivethe“right”optimiza

8、tioncriterion.ThisisdoneinalearningapproachTheseclassicalprinciplesofgroupinghaveinspiredmanyusingadatabaseofhumansegmentedimages.previousapproachestosegmentation.However,theGestaltWeformulatethecomputationalproblemofsegmenta-pri