二維視覺對象分割.pdf

1、浙江大學博士學位論文二維視覺對象分割姓名：趙明申請學位級別：博士專業(yè)：計算機應用指導教師：陳純20040501AbstractIncomputervision，2DvisualobjectsegmentstionisakeyproblemInthisthesis，wefocusedoDtile2Dvisualobjectsegmentationforthetwo21)inputdataofeolnputervison；imageandv

2、ideoSoourresearchincludedtwoparts：ImageObjectSegmentationandVideoObjectSegmentation?！狪mageObject—SegmentationWeusedactivestatisticalobjectmodel，includingActiveShapeModels(ASMjandActiveAppearanceModels(AAM)，forfaceimageob

3、jectseglnentalionp昭proposedweightedASMshapesnbspaceoptimizedASMandsubspaeeoptimizedAAM—ActiveShapeModelsusetheorthogonalprojectiontotileshapesubspaceneglectingtileinformationdmingtilesearchprocedureSo、aWeightedActiveShap

4、eiVlodels(WASM)isproposedusingthesereohinfornlationThesearchinformationisbasedonthelocalappearancemodelofASMtodeterminehowwellthesearchirlgshapem0tchmodelsderivedfionlthetrainingsetItisusedtoprojecttheintermediatesearchs

5、hapetotheshapesubspar：eComparedwithASMsorthogonalprojection，theweightedprojectionCalldragthesearchoutofloeaIminimatobeIIloreaccurateandmorerobustExperimentshavebeendonetoshowtheabilityofthismethodtoalignshapes—ActiveShap

6、eModels(ASM)iscomposedoftwoparts：theASMshapesubspacemodelandtileASMseatchWhilethesetwopartsarecloselyinterrelatedandtheperformanceofASMdependsonbothofthem，existingeffortstreatedthenlseparatelyandhadnotconsideredllowtoopt

7、imizethemoverallIntheirmethodstheASMshapesubspaceumdclwastrainedwithmltcoImideringthesearchprocedurean(1theASMsearchwaaperformedusingthisshapesubspacemodelwithoutconsideringhowitwastrainedSoweproposedanapproachtooptimize

8、theshapemodelwhileconsideringtheASIVIsearehWefirstperforillanerroranalysisofASIVl，andthen“)minimizetheASMerrorweproposeanapprnachwhichoptimizestheshapemodelaccordingtotheASMsearchFortileAShterroranalysis，wedecomposedtheA

9、SMerrorintotwopartswhichareintroducedbytheASMmodelandtheASMsearchrespectivelyWiththisdecompositionweprovethattheoptimalresultsofASMcanbeachievedonlybyoptimizingbothofthemjointlyratherthanseparatelynlrthermorebasedOilthis

10、errordecompositionwedevelopamethodtofindtheoptimalASMshapemodelaccordingtotileASMsearchbyconsideringboththetwodecomposederrorsExperimentalresultsdemonstratethatourmethodcanfindLlmoptimalASMshapemodelrapidlyandimprovethep

11、erformanceofASMsignificantIy一SimilartoAsMActiveAppearanceModels(AAMlisalsocomposedoftwoparts：theAAMsubspacenlodelandtheAAMseal“etl_WhilethesetWOpartsarecloselycorrelatedexistingeffortstreatedtheinseparatelyandhadnotCOILS

12、ldere(1hoartoopthnizethemoverallInthispaper，anapproachisproposedtooptimizetilesubspacemodelwhileconsideringthesearchprocedureWefirstperforlnasubspaceerroranalysis，andthentominimizetheAAMerrorweproposeanapproachwhichoptim

13、izesthesubspacemodelaccordingtothesearchprocedureForthesubspaceerroranMysis，wedecomposedtilesubspaceerrorintotwoparts，whichareintroducedbytilesnbspacemodelandthesearchprocedurerespectivelyThisdecompositionshowsthattheopt