資源描述:
《a multiscale image representation using hierarchicalnew》由會員上傳分享,免費在線閱讀,更多相關(guān)內(nèi)容在教育資源-天天文庫。
1、MULTISCALEMODEL.SIMUL.c2004SocietyforIndustrialandAppliedMathematicsVol.2,No.4,pp.554–579AMULTISCALEIMAGEREPRESENTATIONUSINGHIERARCHICAL(BV,L2)DECOMPOSITIONS?EITANTADMOR?,SUZANNENEZZAR?,ANDLUMINITAVESE?Abstract.Weproposeanewmultiscaleimagedecompositionwhicho?ersahierarchical,a
2、daptiverepresentationforthedi?erentfeaturesingeneralimages.Thestartingpointisavari-ationaldecompositionofanimage,f=u0+v0,where[u0,v0]istheminimizerofaJ-functional,pJ(f,λ0;X,Y)=infu+v=fuX+λ0vY.Suchminimizersarestandardtoolsforimagema-nipulations(e.g.,denoising,deblurring,c
3、ompression);see,forexample,[M.MumfordandJ.Shah,ProceedingsoftheIEEEComputerVisionPatternRecognitionConference,SanFrancisco,CA,1985]and[L.Rudin,S.Osher,andE.Fatemi,Phys.D,60(1992),pp.259–268].Here,u0shouldcapture“essentialfeatures”offwhicharetobeseparatedfromthespuriouscomponent
4、sabsorbedbyv0,andλ0isa?xedthresholdwhichdictatesseparationofscales.Toproceed,weiteratethere?nementstep[u,v]=arginfJ(v,λjkj+1j+1j02),leadingtothehierarchicaldecomposition,f=j=0uj+vk.Wefocusourattentionontheparticularcaseof(X,Y)=(BV,L2)decomposition.Theresultinghierarchicaldeco
5、mposition,f~juj,isessentiallynonlinear.Thequestionsofconvergence,energydecomposition,localization,andadaptivityarediscussed.Thedecompositionisconstructedbynu-mericalsolutionofsuccessiveEuler–Lagrangeequations.Numericalresultsillustrateapplicationsofthenewdecompositiontosyntheti
6、candrealimages.Bothgreyscaleandcolorimagesareconsidered.Keywords.naturalimages,multiscaleexpansion,totalvariation,localization,adaptivityAMSsubjectclassi?cations.26B30,65C20,68U10DOI.10.1137/0306004481.Introductionandmotivations.ImagescouldberealizedasgeneralL2ob-jects,f∈L2(R2)
7、,representingthegreyscaleoftheobservedimage.Likewise,colorimagesaretypicallyrealizedintermsofvector-valuedfunctions,f=(f1,f2,f3)∈L2(R2)3,representingtheRGB-colorscales.Inpractice,themorenoticeablefeaturesofimagesareidenti?edwithinapropersubclassofallL2objects.Mostnoticeablearet
8、heedgesofanimage,whichareknowntobewellquanti?edwithint