Separating a foreground object from the background in a static image involves determining both full and partial pixel coverages, also known as extracting a matte. Previous approaches require the input image to be pre-segmented into three regions: foregroun
AnIterativeOptimizationApproachforUni edImageSegmentationand
Matting
JueWang
Dept.ofElectricalEngineeringUniversityofWashingtonSeattle,WA98195,USAjuew@ee.washington.eduAbstract
Separatingaforegroundobjectfromthebackgroundinastaticimageinvolvesdeterminingbothfullandpartialpixelcoverages,alsoknownasextractingamatte.Previousap-proachesrequiretheinputimagetobepre-segmentedintothreeregions:foreground,backgroundandunknown,whichiscalledatrimap.Partialopacityvaluesarethencom-putedonlyforpixelsinsidetheunknownregion.Thispre-segmentationbasedapproachfailsforimageswithlargeportionsofsemi-transparentforegroundwherethetrimapisdif culttocreateevenmanually.Inthispaperwecombinethesegmentationandmattingproblemtogetherandproposeauni edoptimizationapproachbasedonBeliefPropaga-tion.Weiterativelyestimatetheopacityvalueforeverypixelintheimage,basedonasmallsampleofforegroundandbackgroundpixelsmarkedbytheuser.Experimentalresultsshowthatcomparedwithpreviousapproaches,ourmethodismoreef cienttoextracthighqualitymattesforforegroundswithsigni cantsemi-transparentregions.
1.Introduction
Imagemattingreferstotheproblemofestimatinganopacity(alphavalue)andforegroundandbackgroundcol-orsforeachpixelintheimage.Speci cally,theobservedimageI(z)(z=(x,y))ismodeledasalinearcombinationofforegroundimageF(z)andbackgroundimageB(z)byanalphamap:
I(z)=αzF(z)+(1 αz)B(z)
(1)
whereαzcanbeanyvaluein[0,1].Ifweconstrainthealphavaluetobeeither0or1,thenthemattingproblemde-gradestobethesegmentationproblem,whereeachpixelisassignedtobeeitherfullyforegroundorfullybackground.
MichaelF.CohenMicrosoftResearchRedmond,WA98052,USAmcohen@microsoft.com
Fornaturalimages,allvaluesα,FandBneedtobeestimatedforeverypixelthusitisinherentlyanunder-constrainedproblem.Previousapproachesandtheonere-portedheredependonuserinteractioncombinedwithpriorassumptionsonimagestatisticstoobtaingoodmattes.Asinearlierworkwewillrequireminimaluserinput.Incontrasttopreviousmethodswewillnot rstsegmenttheimageintoforeground,background,andunknownregions.Suchmethodsfailinimageswithsigni cantpartialpixelcoverageandforforegroundobjectscontainingmanyholes.Rather,wecallonabeliefpropagationoptimizationmethodtodirectlydetermineαvalues.
1.1.PreviousWork
Earlymattingapproachestrytosimplifytheproblembyphotographingforegroundobjectsagainstaconstant-coloredbackground,whichiscalledbluescreenmatting[1].However,evenknowingthebackgroundcolorisstillinsuf cienttofullyconstraintheproblemthussomesim-pleconstraintsaremade,whichrequireexpertlevelpa-rametertuningandcanfailonfairlysimpleforegrounds.Mishima[8]improvesthistechniquebyintroducingstatis-ticalmethodsbasedonrepresentativeforegroundandback-groundsamples.
Recentapproachesattempttoextractforegroundmattesdirectlyfromnaturalimages.ThemostsuccessfulsystemsincludeKnockout2[3],theapproachproposedbyRuzonandTomasi[7],Bayesianmatting[12]andPoissonmatting[5].Allthesesystemsstartbyhavingtheusersegmenttheimageintothreeregions:de nitelyforeground,de nitelybackgroundandunknown,whichisoftenreferredtoasatrimap.TheproblemisthusreducedtoestimatingF,Bandαintheunknownregion.TheKnockoutsystemex-trapolatesknownforegroundandbackgroundcolorsintotheunknownregionandestimatesαsaccordingtothem.RuzonandTomasiarethe rsttotakeaprobabilisticviewoftheproblem.Theyanalyzeforegroundandbackground
热门文档