aliasing due to poor quantization is suppressed without altering the general topography of the surface (both pictures are flat-shaded). In this paper, we present an efficient way to denoise bivariate data like height fields, color pictures or vector fields
AnisotropicFeature-PreservingDenoising
ofHeightFieldsandBivariateData
MathieuDesbrun MarkMeyer PeterSchr¨oder AlanH.Barr
Caltech-


USC
(a)(b)
Figure1:Marselevationmap:(a)rawdata,(b)smoothversionafteranisotropicdiffusion.Noticehow,withournon-uniformdiffusion,the
aliasingduetopoorquantizationissuppressedwithoutalteringthegeneraltopographyofthesurface(bothpicturesare at-shaded).
Abstract
Inthispaper,wepresentanef cientwaytodenoisebivariatedatalikeheight elds,colorpicturesorvector elds,whilepreservingedgesandotherfeatures.Mixingsurfaceareaminimization,graph ow,andnonlinearedge-preservationmetrics,ourmethodgeneral-izespreviousanisotropicdiffusionapproachesinimageprocessing,andisapplicabletodataofarbitrarydimension.Anothernotabledifferenceistheuseofamorerobustdiscretedifferentialoperator,whichcapturesthefundamentalsurfaceproperties.Wedemonstratethemethodonrangeimagesandheight elds,aswellasgreyscaleorcolorimages.
CRCategories:I.3.7[ComputerGraphics]Three-DimensionalGraphicsandRealism;I.4.3[ImageProcessingandComputerVi-sion]Enhancement.
Keywords:Anisotropicdiffusion,Re-parameterization,Surface ows,Edgepreservation,Imageprocessing,Surfacefairing.
whilereducingnoiseinthesignal.
Theconceptofnoiseremovalwithpreservationofedgescanbeusedfordenoisingtexturesacquiredfromimages,orgeneralizedformeshestopreservefeatureswhileremovingsmallbumps,orevenforreducingquantizationeffectsinheight elds.Inthispaper,weproposeatechniqueforalltheseapplications.Usingrobustdif-ferentialgeometrytoolsdevelopedinthediscretedomainforcom-putergraphics,wede neanon-linearanisotropicdiffusionoperatorvalidforanybivariatedata(height elds,greyscaleimages,colorimages,tensorimages).Weexplaintherelationsbetweenourap-proachandpreviousworkinimageprocessing,andshowresultstodemonstrateitsreliability.
1.1AnisotropicDiffusioninImageProcessing
The rstinhomogeneousdiffusionmodelwasintroducedbyPer-onaandMalik[PM90].Theideawastovarytheconductionspa-tiallytofavornoiseremovalinnearlyhomogeneousregionswhileavoidinganyalterationofthesignalalongsigni cantdiscontinu-ities(see[TT99]foranintuitiveexplanationofthistechnique).ThechangeinintensityIovertimewasde nedas:
It=div(g(|| I||) I)with:g(x)=
11+
x2α
1Introduction
Theproblemofsmoothingsurfacesincomputergraphicsiscloselyrelatedtothesmoothingofimagesincomputervision.Inbothcases,wewishtosmootha2-dimensionaldataset,whilepreserv-ingimportantfeaturessuchasdiscontinuities.Thediscontinuitiesrepresentcliffsinheight elds,andedgesinimages(edgesinim-agesarisewhentheassociatedheight- eldvalues—intensitiesoftheimage—changesharplywithposition).
Toreducethenoiseinimages,earlyresearchhasadvocatedtheuseofthelaplacianasalocaldifferentialoperator.Diffusingthesignalusinglaplaciansmoothingwillreducehighfrequencynoise.Unfortunately,anunintendedconsequenceisthatthenoiseisdif-fuseduniformlyinscreenspace.Sharpedgesandotherfundamen-talfeaturesofanimagearethenlost,blurredawaybytheuniformdiffusion.Consequently,anisotropicoperatorshavebeenproposed.Theycandiffusethesignalnon-uniformlytobetterpreserveedges,
.(1)
Manydifferentvariationsontheconductionfunctionghavebeenproposed[ROF92,ABBFC97,ALM92],andrecentlyahigheror-derPDEhasbeenintroducedbyTumblin[TT99]inthecontextofdisplayinghighcontrastcomputergraphicspictures.Similartechniqueshavebeenusedtovisualizecomplex ow elds,asin[PR99].Alloftheseapproachesrelyonisophotesoftheim-age(seeFigure2(a)):theanisotropicdiffusionequationcanbein-terpretedasadiffusionmainlyinthedirectiontangentialtoeachisophote.Therefore,discontinuitiespresentintheorthogonaldirec-tionarenotlost,asexplainedin[KDA97].Typically, nitediffer-enceschemesareusedtodiscretizethedifferentialoperatorsused.
bivariate polynomial surfaces [2], or other higher...Hu, G., Peng, Q.: Anisotropic denoising of ...Non-iterative, feature preserving mesh smoothing. ...
bivariate polynomial surfaces [2], or other higher...Hu, G., Peng, Q.: Anisotropic denoising of ...Non-iterative, feature preserving mesh smoothing. ...
热门文档