ReceivedSeptember82009acceptedinrevisedversionSeptember72010CommunicatedbyMartinBurgeryInstituteforNumericalandAppliedMathematicsGeorgAugustUniversityofGottingenLotzestr1618D37083Go ID: 203173
Download Pdf The PPT/PDF document "COMMUN.ATH.SCI.c\r2011InternationalPress..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
COMMUN.ATH.SCI.c\r2011InternationalPressVol.9,No.2,pp.413{457UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGCAROLA-BIBIANESCHONLIEByANDANDREABERTOZZIzAbstract.Higherorderequations,whenappliedtoimageinpainting,havecertainadvantagesoversecondorderequations,suchascontinuationofbothedgeandintensityinformationoverlargerdistances.Discretizingafourthorderevolutionequationwithabruteforcemethodmayrestrictthetimestepstoasizeuptoorderx4wherexdenotesthestepsizeofthespatialgrid.Inthisworkwepresentecientsemi-implicitschemesthatareguaranteedtobeunconditionallystable.WeexplainthemainideaoftheseschemesandpresentapplicationsinimageprocessingforinpaintingwiththeCahn-Hilliardequation,TV-H 1inpainting,andinpaintingwithLCIS(lowcurvatureimagesimpliers).Keywords.Imageinpainting,higherorderequations,numericalschemes.AMSsubjectclassications.35G25;34K28.1.IntroductionAnimportanttaskinimageprocessingistheprocessofllinginmissingpartsofdamagedimagesbasedontheinformationgleanedfromthesurroundingareas.Itisessentiallyatypeofinterpolationandiscalledinpainting.Therebyonecouldrestoreimageswithdamagedpartsdueto,forinstance,intentionalscratching,aging,orweather.Oronecanrecoverobjectswhichareoccludedbyotherobjects,wherewithinthiscontexttheprocessiscalleddisocclusion.Infacttheapplicationsofimageinpaintingarecountless.Fromtherestorationofancientfrescoes[3],tothemedicalneedsofreducingartifactsinMRI-,CT-orPETimagingreconstructions[47],digitalimageinpaintingisubiquitousinourmoderncomputerizedsociety.SincetherstworksonimageinpaintingbyMumford,NitzbergandShiota[57],MasnouandMorel[52],Caselles,Morel,SbertandGillette[21],andBertalmioetal[10],mucheorthasgoneintodevelopingdigitalalgorithms.Thesemethodsincludethetexturesynthesisandexemplar-basedapproach(see,e.g.,[20,29,32,72])andanumberofvariational-andPDE-basedapproaches.Thispaperfocusesonthelatter.Inmathematicalterms,imageinpaintingcanbedescribedinthefollowingway:letfbethegivenimagedenedonanimagedomain\n.TheproblemistoreconstructtheoriginalimageuinthedamageddomainD\n,calledtheinpaintingdomain.Moreprecisely,let\nR2beanopenandboundeddomainwithLipschitzboundary,B1;B2twoBanachspacesandf2B1bethegivenimage.Ageneralvariationalapproachininpaintingcanbewrittenasmin2B2nE(u)=R(u)+k(f u)k2B1o;(1.1)whereR:B2Rand(x)=(0\nnD0D;(1.2) Received:September8,2009;accepted(inrevisedversion):September7,2010.CommunicatedbyMartinBurger.yInstituteforNumericalandAppliedMathematics,Georg-AugustUniversityofGottingen,Lotzestr.16-18,D-37083Gottingen,Germany(c.schoenlieb@math.uni-goettingen.de).zDepartmentofMathematics,UCLA(UniversityofCaliforniaLosAngeles),520PortolaPlaza,LosAngeles,CA90095-1555,USA(bertozzi@math.ucla.edu).413 414UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGisthecharacteristicfunctionof\nnDmultipliedbyaconstant01.R(u)denotestheregularizingtermandk(f u)kB1thesocalleddelitytermoftheinpaintingapproach.B2B1ingeneral,signifyingthesmoothingeectoftheregularizingtermontheminimizeru2B2.DependingonthechoiceoftheregularizingtermRandtheBanachspacesB1,B2,variousinpaintingapproacheshavebeendeveloped.Themostfamousmodelisthetotalvariation(TV)model,whereR(u)=R\njrujdxdenotesthetotalvariationofu,B1=L2(\n)andB2=BV(\n)thespaceoffunctionsofboundedvariation;cf.[23,25,61,60].AvariationalmodelwitharegularizingtermcontaininghigherorderderivativesistheEulerselasticamodel[26,27,52]whereR(u)=R\n(a+b2)jrujdxwithpositiveweightsaandb,andcurvature=r(ru=jruj).Otherexamplestobementionedfor(1.1)aretheactivecontourmodelbasedonMumfordandShahssegmentation[68],theinpaintingschemebasedontheMumford-Shah-Eulerimagemodel[35],inpaintingwiththeNavier-Stokesequation[11],andwavelet-basedinpainting[28,30],onlytogivearoughoverview.ForamorecompleteintroductiontoimageinpaintingusingPDEswereferto[26,18,63].1.1.Second-versushigher-orderinpaintingapproaches.Secondordervariationalinpaintingmethods(wheretheorderofthemethodisdeterminedbythederivativesofhighestorderinthecorrespondingEuler-Lagrangeequation),likeTVinpainting,havedrawbacksasintheconnectionofedgesoverlargedistances(Con-nectivityPrinciple,cf.Figure1.1)andthesmoothpropagationoflevellines(setsofimagepointswithconstantgrayvalue)intothedamageddomain(CurvaturePreser-vation,cf.Figure1.2).Thisisduetothepenalizationofthelengthofthelevellineswithintheminimizingprocesswithasecondorderregularizer,connectinglevellinesfromtheboundaryoftheinpaintingdomainviatheshortestdistance(linearinterpo-lation).TheregularizingtermR(u)=R\njrujdxintheTVinpaintingapproach,forexample,canbeinterpretedviathecoareaformula,whichgivesminZ\njrujdx()min Z1 1length( )d;where =fx2\n:u(x)=gisthelevellineforthegrayvalue.IfweconsiderontheotherhandtheregularizingtermintheEulerselasticainpaintingapproachthecoareaformulareadsminZ\n(a+b2)jrujdx()min Z1 1alength( )+bcurvature2( )d:(1.3)Thusnotonlythelengthofthelevellinesbutalsotheircurvatureispenalized(wherethepenalizationofeachdependsontheratiob=a).Thisresultsinasmoothcon-tinuationoflevellinesovertheinpaintingdomainalsooverlargedistances;compareFigures1.1and1.2.Theperformanceofhigherorderinpaintingmethodscanalsobeinterpretedviathesecondboundarycondition,whichisnecessaryforthewell-posednessofthecorrespondingEuler-Lagrangeequationoffourthorder.Notonlyarethegrayvaluesoftheimagespeciedontheboundaryoftheinpaintingdomain,butalsothegradientoftheimagefunction,namelythedirectionofthelevellines,isgiven.Inanattempttosolveboththeconnectivityprincipleandthestaircasingeectresultingfromsecondorderimagediusions,anumberofthirdandfourthorderdiusionshavebeensuggestedforimageinpainting.TherstworkconnectingimageinpaintingtoathirdorderPDE(partialdierentialequation)isthetransportprocess C.B.SCHONLIEBANDA.BERTOZZI415 Fig.1.1.TwoexamplesofcurvaturebasedinpaintingcomparedwithTVinpaintingfrom[26].InthecaseoflargeaspectratiostheTVinpaintingfailstocomplytotheConnectivityPrinciple. Fig.1.2.AnexampleofelasticainpaintingcomparedwithTVinpaintingfrom[27].Despitethepresenceofhighcurvature,TVinpaintingtruncatesthecircleinsidetheinpaintingdomain(linearinterpolationoflevellines,i.e.,CurvaturePreservation).DependingontheweightsaandbEulerselasticainpaintingreturnsasmoothlyrestoredobject,takingthecurvatureofthecircleintoaccount.ofBertalmioetal[10].Theimageinformation,modeledbyu,istransportedintotheinpaintingdomainalongthelevellinesoftheimage.TheresultingschemeisadiscretemodelbasedonthenonlinearPDEu=r?uru;andissolvedinsidetheinpaintingdomainDusingtheimageinformationfromasmallstripearoundtheboundaryofD.Theoperatorr?denotestheperpendiculargradient( @y;@x).Duetothelackofcommunicationamongthelevellines,thetransportationmayresultinkinksorcontradictionsinsidetheinpaintingdomain.Thusin[10]theequationaboveisimplementedwithintermediatestepsofanisotropicdiusion.In[11]theauthorsdevelopatheoryfortheproperboundaryconditionsin[10]bymakingaconnectiontotheNavier-Stokesequations.Thetwoconditionsonthe\boundary"oftheinpaintingdomaincorrespondtothenoslipconditionforNavier-Stokes.AvariationalthirdorderapproachtoimageinpaintingisCDD(CurvatureDrivenDiusion)[24].Tosolvetheproblemofconnectinglevellinesalsooverlarge 416UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGdistances(connectivityprinciple),thelevellinesarestillinterpolatedlinearly.Thedrawbacksofthethird-orderinpaintingmodels[10]and[24]havedrivenChan,KangandShen[27]toareinvestigationoftheearlierproposalofMasnouandMorel[52]onimageinterpolationbasedonEulerselasticaenergy(1.3).ThefourthorderelasticainpaintingPDEcombinesCDD[24]andthetransportprocessofBertalmioetal[10],andisabletosolveboththeconnectivityprincipleandthestaircasingeect.OtherrecentlyproposedhigherorderinpaintingalgorithmsareinpaintingwiththeCahn-Hilliardequation[13,14],TV-H 1inpainting[19,64]andcombinationsofsecondandhigherordermethods,e.g.[51].Inthispaperweareespeciallyinterestedinthree,rathernew,fourth-orderin-paintingschemes.Namely,weshalldiscussCahn-Hilliardinpainting,TV-H 1inpaint-ing,andinpaintingwithLCIS(lowcurvatureimagesimpliers).Westartthediscus-sionwiththeinpaintingofbinaryimagesusingtheCahn-Hilliardequation[13,14].Theinpaintedversionuoff2L2(\n)isconstructedbyfollowingtheevolutionofu= u+1 F0(u)+(f u);(1.4)whereF(u)isasocalleddouble-wellpotential,e.g.,F(u)=u2(u 1)2.Theapplica-bilityoftheCahn-HilliardequationfortheinpaintingofbinaryimagesisduetothedoublewellpotentialF(u)intheequation.Thetwowellscorrespondtovaluesofuthataretakenbymostofthegrayscalevalues.Choosingapotentialwithwellsatthevalues0(black)and1(white),Equation(1.4)thereforeprovidesasimplemodelfortheinpaintingofbinaryimages.Theparameterdeterminesthesteepnessofthetransitionbetween0and1.Further,thefourthorderregularizingtermintheequa-tionprovidestheadvantagesofhigherorderinpaintingapproacheswhichhavebeendiscussedbefore,suchastheabilitytoconnectlevellinesalsooverlargedistances(cf.(1.3)).ThesecondmethodofinterestinthispaperisageneralizationoftheCahn-Hilliardinpaintingapproachtograyvalueimageswhichhasbeenrecentlyproposedin[19,64]andiscalledTV-H 1inpainting.Thereintheinpaintedimageuoff2L2(\n)shallevolveviau=p+(f u);p2@TV(u);(1.5)withTV(u)=(R\njrujdx;ifju(x)j1a.e.in\n;+;otherwise;where@TV(u)denotesthesubdierentialofthefunctionalTV(u).TobuildtheconnectiontoCahn-Hilliardinpaintingtheauthorsin[19]showthatsolutionsofanappropriatetime-discreteCahn-Hilliardinpaintingapproach -converge,as0,tosolutionsofanoptimizationproblemregularizedwiththeTVnorm.Asimilarformofthisapproachappearsinthecontextofdecompositionandrestorationofgrayvalueimages;seeforexample[49,58,70].Further,inBertalmioetal[12],anapplicationofthemodelfrom[70]toimageinpaintingisproposed.Incontrasttotheinpaintingapproach(1.5)theauthorsin[12]useamoregeneralformoftheTV-H 1approachforadecompositionoftheimageintocartoonandtexturepriortotheinpaintingprocess.Thelatterisaccomplishedwiththemethodpresentedin[10].Moreover,we C.B.SCHONLIEBANDA.BERTOZZI417wouldliketomentionthatin[45]theauthorsconsideracomplexGinzburg-Landauenergyforinpaintingofgrayscale-andcolorimages.ThethirdinpaintingmodelwearegoingtodiscussisinpaintingwithLCIS(LowCurvatureImageSimplier).ThishigherorderinpaintingmodelismotivatedbytwofamoussecondordernonlinearPDEsinimageprocessing|theworksofRudin,OsherandFatemi[60]andPeronaMalik[59].Thesemethodsarebasedonanonlinearversionoftheheatequationu=r(g(jruj)ru);inwhichgissmallinregionsofsharpgradients.LCISrepresentafourthorderrelativeofthesenonlinearsecondorderapproaches.Theyareproposedin[69]andlaterusedbyBertozziandGreerin[15]forthedenoisingofpiecewiselinearsignals.InthispaperweconsiderLCISforimageinpainting.Withf2L2(\n)ourinpaintedimageuevolvesintimeasu= r(g(u)ru)+(f u);withthresholdingfunctiong(s)=1 1+s2.Notethatwithg(u)ru=r(arctan(u))theaboveequationcanberewrittenasu= (arctan(u))+(f u):(1.6)1.2.Numericalsolutionofhigher-orderinpaintingequations.Onemainchallengeininpaintingwithhigherorder\rowsistheireectivenumeri-calimplementation.Discretizingafourthorderevolutionequationwithabrute-forcemethodmayrestrictthetimestepstoasizeuptoorderx4wherexdenotesthestepsizeofthespatialgrid.Suchabrute-forcemethodiscomputationallyprohibitiveandhenceitisessentiallyneverdone;see,e.g.,[65].Thenumericalsolutionofhigher-orderequations,likethinlms,phaseeldmod-els,surfacediusionequations,andmanymore,occupiedabigpartofresearchinnumericalanalysisinthelastdecades.In[31]theauthorsproposeasemiimplicit-nitedierenceschemeforthesolutionofsecondorderparabolicequations.Adiusiontermisaddedimplicitlyandsubtractedexplicitlyintimetothenumericalschemeinordertosuppressunstablemodes.Smerekausesthisideatosolvethefourth-ordersurfacediusionequation;cf.[65].ThesameideaisappliedbyGlasnertoaphaseeldapproachfortheHele-Shawinterfacemodel;cf.[40].Besidesthenitedier-enceapproximations,therealsoexistmanyniteelementalgorithmsforfourth-orderequations.Barrett,Blowey,andGarckepublishedaseriesofpapersonthesolutionofvariousCahn-Hilliardequations;cf.[5,6,7].ForthesharpinterfacelimitofCahn-Hilliard,i.e.,theHele-Shawmodel,FengandProhlanalyzeniteelementmethodsin[37,38].Finiteelementmethodsforthinlmequationsarestudied,forinstance,in[8,46].Forimageinpainting,ecientnumericalschemesforhigher-ordermethodsisanactiveareaofresearch.Asdiscussedin[26]oneofthemostinterestingopenproblemsindigitalinpaintingis,infact,thefastandexactdigitalrealization.InthecaseofCahn-Hilliardinpainting,in[13]theauthorsproposeasemi-implicitschemewhichconstitutesthecommonnumericalmethoddiscussedinthispaper.Theyverifyitscomputationalsuperioritycomparedwithcurrentlyusednumericalmethodsforthreecurvaturedrivenapproaches.ItturnsoutthatCahn-Hilliardinpaintingperformsatleastoneorderofmagnitudefasterthanthecurvaturemethods.In[33,34]Elliott 418UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGandSmithemanproposeaniteelementmethodforTV-H 1minimizationinthecontextofimagedenoisingandcartoon/texturedecomposition.Theyalsoproverig-orousresultsabouttheapproximationandconvergencepropertiesoftheirscheme.AnextensionoftheirapproachtoTV-H 1inpaintingwouldbeinteresting.Notethat,however,thedierenceoftheinpaintingapproachfromdenoisinganddecom-positionisthattheformerdoesnotfollowavariationalprincipleandthedelitytermislocallydependentonthespatialposition.AnotheralgorithmforTV-H 1inpaintingisproposedbyoneoftheauthorsin[62].ThisworkgeneralizesthedualapproachofChambolle[22]andBectetal.[9]fromanL2delitytermtoanH 1delityandextendsitsapplicationfromTV-H 1denoising[1,2]toimageinpainting.Themainmotivationfortheworkin[62]isthatwiththeproposedalgorithmthedomaindecompositionapproachdevelopedin[39]canbeappliedtothehigher-ordertotalvariationcase.BeingabletoapplydomaindecompositionmethodstoTV-H 1inpaintingcanresultinatremendousaccelerationofcomputationalspeedduetotheabilitytoparallelizethecomputation.Anotherveryrecentapproachinthisdirectionis[18],wheretheauthorsproposeamultigridapproachforinpaintingwithCDD.InthispaperwediscussanecientsemiimplicitapproachbasedonanumericalmethodpresentedinEyre[36](alsocf.[71])calledconvexitysplitting.Convexitysplittingwasoriginallyproposedtosolveenergyminimizingequations.Weconsiderthefollowingproblem:LetE2C2(RN;R)beasmoothfunctionalfromRNintoR,whereNisthedimensionofthedataspace.Let\nbethespatialdomainofthedataspace.Findu2RNsuchthat(u= rE(u);in\n;u(:;t=0)=u0;in\n;(1.7)withinitialconditionu02RN.ThebasicideaofconvexitysplittingistosplitthefunctionalEintoaconvexandaconcavepart.Inthesemiimplicitscheme,thecon-vexpartistreatedimplicitlyandtheconcaveoneexplicitlyintime.Underadditionalassumptionson(1.7),thisdiscretizationapproachisunconditionallystable,consis-tent,andrelativelyeasytoapplytoalargerangeofvariationalproblems.Moreoverweshallseethattheideaofconvexitysplittingcanbeappliedtomoregeneralevo-lutionequations,andinparticulartothosethatdonotfollowavariationalprinciple,especiallytotheinpaintingEquations(1.4)and(1.5).Convexitysplittingmethods,althoughpossiblynotunderthesamename,alreadyhavealongtraditioninseveralpartsofnumericalanalysis.Inniteelementapprox-imationsforPDEs,examplesforsuchnumericalschemescanbefoundintheworksofBarrett,Blowley,andGarcke;cf.[4]Equation(3.42)foranapplicationtoamodelforphaseseparation.In[35]anitedierenceschemeforsecond-orderparabolicequationsispresentedwhichalsousestheconvexitysplittingidea;cf.Equation(5.4)in[35].Furtherconvexitysplittingisalsodiscussedinamoregeneraloptimizationcontext;cf.[73]Chaptertwoforanoverviewonthistopic.Themainpartofthepaperillustratestheapplicationoftheconvexitysplittingideatothethreefourth-orderinpaintingapproaches(1.4),(1.5),and(1.6).Moti-vatedbytheanalysisin[17],weshowthatwiththisnumericalapproachweareableto(approximately)computestrongsolutionsofthecontinuousproblemwithanunconditionallystablenitedierencescheme.Thenumericalschemeissaidtobeunconditionallystableifallsolutionsofthedierenceequationareboundedindepen-dentlyfromthetimestepsize;cf.Denition2.2.Moreover,weproveconsistencyoftheseschemesandconvergencetotheexactsolution.Further,wepresentnu- C.B.SCHONLIEBANDA.BERTOZZI419mericalresultsdemonstratingtheeectofthehigherorderregularizingtermintheapproaches.InthecaseofTV-H 1inpaintingandinpaintingwithLCISwedirectlycomparethevisualresultswiththesecondorderTVinpaintingmethod.Organizationofthepaper.InSection2theideaofconvexitysplittingispre-sented.AfteranintroductiontogradientsystemswestateandproveEyre'stheoremabouttheunconditionalstabilityoftheconvexitysplittingscheme.Sections3-5arededicatedtotheapplicationofconvexitysplittingtoCahn-Hilliardinpainting(1.4),TV-H 1inpainting(1.5),andinpaintingwithLCIS(1.6).InthecaseofCahn-HilliardandTV-H 1inpaintingthecorrespondingEquations(1.4)and(1.5)arenotstrictlygradient\rows,buttheirevolutionisthesumofthegradientsoftwodierentener-gies.Here,convexitysplittingisappliedtoeachoftheseenergiesandresultsinasemi-implicitschemeforthewholeevolution.Rigorousproofsfortheconsistencyofthenumericalscheme,theboundednessofnumericalsolutionsandtheirconvergencetotheexactsolutionaregiven.Foreachoftheseinpaintingalgorithmsnumericalresultsarepresented.Intheconclusionofthepaperopenproblemsarediscussed.Notation.Inthispaperwediscussthenumericalsolutionofevolutionarydif-ferentialequations.ThereforewehavetodistinguishbetweentheexactsolutionuofthecontinuousequationandtheapproximativesolutionUofthecorrespondingtimediscretenumericalscheme.WewritecapitalUkforthekthsolutionofthediscreteequationandsmalluk=u(kt)forasolutionofthecontinuousinpaintingequationattimektwithtimestepsizet.Letekdenotethetemporaldiscretizationerrorgivenbyek=uk Uk.Insubsectiontwo,uandUarevectorsinRN,whereNdenotesthedimensionofthedata.InallotherpartsofthispaperuandUareassumedtobeelementsinL2(\n).LetE2C2(H;R)denoteafunctionalfromasuitableHilbertspaceHtoR,andrE(u)itsrstvariationwithrespecttou.Inthediscreteset-tingH=RN.ThroughoutthispaperkkdenotesthenorminL2(\n)(ortheEuclideannorminthediscretesetting),andh;itheinnerproductinL2(\n)(orinRNinthedis-cretesetting).Finally,sinceweposeallthreeinpaintingapproaches(1.4)-(1.6)withNeumannboundaryconditions,wehavetodenethenon-standardspaceH 1(\n)asH 1(\n)=nF2H1(\n)jhF;1i(H1);H1=0o;withnormkk 1:=\r\rr 1\r\rL2(\n).Therebytheoperator 1denotesthein-verseofwithNeumannboundaryconditions.Inmoredetail,letH1(\n):= 2H1(\n):R\n dx=0 .Thenu= 1F2H1(\n)istheuniqueweaksolutionofthefollowingproblem:u F=0;in\n;ru=0;on@\n:Foramoreelaboratederivationoftheabovespacewereferto[19],AppendixA.2.TheconvexitysplittingideaAsalreadydiscussedintheIntroduction,convexitysplittingmethodsareusedinawiderangeofoptimizationproblems;cf.Section1.2forrelevantreferences.Orig-inallydesignedtosolvegradientsystems,weshallseeinthispaperthatconvexitysplittingschemesarerelevantformoregeneralproblems,i.e.,forevolutionequationswhichdonotfollowavariationalprinciple.SeeSections3-5forourthreeinpainting 420UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGapproaches(1.4)-(1.6).Firstweintroducethenotionofgradient\rowsandtheapplicationofconvexitysplit-tingmethodsinthiscontext.TodosowefollowtheexplanationsandnotationsinEyre'swork[36].WeconsiderEquation(1.7).IfEfulllsthefollowingconditions;(i)E(u)0;8u2RN;(ii)E(u)!1askuk!1;(iii)hJ(rE)(u)u;ui8u2RN;(2.1)thenEquation(1.7)iscalledagradientsystemanditssolutionsarecalledgradient\rows.TherebyJ(rE)(u)istheJacobianofrEinu,2Randh:;:idenotestheinnerproductonRNwithcorrespondingnormkuk2=hu;ui.Allgradientsystemssatisfythedissipationproperty,i.e.,dE(u) dt= krE(u)k2;andthereforeE(u(t))E(u0)forallt0.IfE(u)isstrictlyconvex,i.e.,incondition(2.1)(iii)ispositive,thenonlyasingleequilibriumforthegradientsystemexists.Unconditionallystableanduniquelysolvablenumericalschemesexistfortheseequations(cf.[66]).IfE(u)isnotstrictlyconvex,i.e.,0,multipleminimizersmayexistandthegradient\rowcanpossiblyexpandinu(t).Thestabilityofanexplicitgradientdescentalgorithm,i.e.,Uk+1=Uk trE(Uk),inthiscasemayrequireextremelysmalltimesteps,dependingofcourseonthefunctionalE.Forfourthorderinpaintingapproaches,forinstance,E(Uk)containssecondorderderivativesresultinginarestrictionoftuptoorder(x)4(wherexdenotesthestepsizeofthespatialdiscretization).Thereforethedevelopmentofstableandecientdiscretizationsfornon-convexfunctionalsEishighlydesirable.ThebasicideaofconvexitysplittingistowritethefunctionalEasE(u)=Ec(u) E(u);(2.2)whereEo2C2(RN;R)andEo(u)isstrictlyconvexforallu2RN;o2fc;eg:(2.3)Thesemi-implicitdiscretizationof(1.7)isthengivenbyUk+1 Uk= t(rEc(Uk+1) rE(Uk));(2.4)whereU0=u0.Remark2.1.WewanttoanticipatethatthesettingofEyre,andhencethesubsequentpresentationofconvexitysplitting,isapurelydiscreteone.Neverthelessitactuallyholdsinamoregeneralframework,i.e.,formoregeneralgradient\rows.InthecaseofanL2gradient\rowforexample,theJacobianJofthediscretefunctionalEjusthastobereplacedbythesecondvariationofthecontinuousfunctionalEinL2(\n).Inthefollowingweshowthatconvexitysplittingcanbeappliedtotheinpaintingapproaches(1.4),(1.5),and(1.6),andproducesunconditionallygradientstableorunconditionallystablenumericalschemes. C.B.SCHONLIEBANDA.BERTOZZI421Definition2.1.[36]Aone-stepnumericalintegrationschemeisunconditionallygradientstableifthereexistsafunctionE(:):RNRsuchthat,forallt0andforallinitialdata:(i)E(U)0forallU2RN,(ii)E(U)!1askUk!1,(iii)E(Uk+1)E(Uk)forallUk2RN,(iv)IfE(Uk)=E(U0)forallk0thenU0isazeroofrEfor(1.7)and(2.1).NotethatCahn-Hilliardinpainting(1.4)andTV-H 1inpainting(1.5)arenotgivenbygradient\rows.Hence,inthecontextoftheseinpaintingmodelsthemeaningofunconditionalstabilityhastoberedened.Namely,inthecaseofanevolutionequationwhichdoesnotfollowagradient\row,acorrespondingdiscretetimesteppingschemeissaidtobeunconditionallystableifsolutionsofthedierenceequationareboundedwithinanitetimeinterval,independentlyofthestepsizet.Definition2.2.LetubeanelementofasuitablefunctionspaceHdenedon\n[0;T],with\nR2openandbounded,andT0.LetfurtherGbearealvaluedfunctionandu=G(u;Du)beapartialdierentialequationwithspacederivativesDu,=1;:::;4.AcorrespondingdiscretetimesteppingmethodUk+1=Uk+tGk(Uk;Uk+1;DUk;DUk+1);(2.5)whereGkisasuitableapproximationofGinUkandUk+1isunconditionallystable,ifallsolutionsof(2.5)areboundedforallt0andallksuchthatktT.consistentiflim!0k(t)=0;wherek(t)isthelocaltruncationerroroftheschemeanddenedask(t)=uk+1 uk t Gk(uk;uk+1;Duk;Duk+1);(2.6)anduk=u(kt)istheexactsolutionattimet=kt.Inwhatfollowsweabbreviatekfork(t).Moreover,wedenetheglobaltruncationerrortobe(t)=maxkkk(t)kH:Anumericalschemeissaidtobeoforderpintimeif(t)=O(tp)fort0:WestartwithatheoremofEyre[36].Theproofpresentedbelowfollowsthesameargumentsasin[36]withadditionaldetails.Theorem2.3([36]Theorem1).LetEsatisfy(2.1),andEcandEsatisfy(2.2)-(2.3).IfE(u)additionallysatiseshJ(rE)(u)u;ui (2.7) 422UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGwhen0in(2.1)(iii),thenforanyinitialconditionthenumericalscheme(2.4)isconsistentwith(1.7),gradientstableforallt-0.8;أ倀0,andpossessesauniquesolutionforeachtimestep.Thelocaltruncationerrorforeachstepisk=t 2(J(rEc(^u))+J(rE(^u)))rE(u());forsome2(kt;(k+1)t)andforsome^uintheparallelopipedwithoppositeverticesatUkandUk+1.Remark2.2.Condition(2.7)inTheorem2.3isequivalenttotherequirementthatalltheeigenvaluesofJ(rE)dominatethelargesteigenvalue of J(rE),i.e.,hJ(rE)(u)u;ui(2.7) (2.1)h J(rE)(u)u;uiforallu2RN,or^jj;foralleigenvalues^0ofE:(2.8)Proof.(Eyre[36]).Theunconditionalgradientstabilityof(2.4)inthesenseofDenition2.1isestablishedrst.Byourassumptionsin(2.1)properties(i)and(ii)inDenition2.1immediatelyfollow.Property(iv)followsfromthegeneralbehaviorofgradientsystems,i.e.,ifE(Uk)=E(U0)forallk0thenU0isan!-limitpointof(1.7)and(2.1)andhenceU0isazeroofrE(cf.[48]).Themainpartoftheproofconsistsofthevericationofproperty(iii).NamelywehavetoshowthatE(Uk+1)E(Uk);8Uk2RN:TodosoweconsiderthedierenceE(Uk+1) E(Uk).TheproofisbyrepeatedapplicationofTaylor'stheorem.WestartwithanexactexpansionofEaboutUk+1uptosecondorderandobtainE(Uk)=E(Uk+1) hrE(Uk+1);Uk+1 Uki+1 2hJ(rE(Uk+1 (Uk+1 Uk)))Uk+1 Uk;Uk+1 Ukiforsome2(0;1).Thenbyassumption(iii)in(2.1)wegetE(Uk+1) E(Uk)hrE(Uk+1);Uk+1 Uki+jjkUk+1 Ukk2:By(2.2)and(2.4)thisisthesameasE(Uk+1) E(Uk)hrEc(Uk+1) rE(Uk+1);Uk+1 Uki+jjkUk+1 Ukk2 1 t(Uk+1 Uk)+rEc(Uk+1) rE(Uk);Uk+1 Uk= hrE(Uk+1) rE(Uk);Uk+1 Uki+jj 1 tkUk+1 Ukk2:(2.9)Similarly,weTaylorexpandEaboutUk+1andUk,respectively,asE(Uk)=E(Uk+1) hrE(Uk+1);Uk+1 Uki+1 2hJ(rE(Uk+1 1(Uk+1 Uk)))Uk+1 Uk;Uk+1 Uki; C.B.SCHONLIEBANDA.BERTOZZI423andE(Uk+1)=E(Uk)+hrE(Uk);Uk+1 Uki+1 2hJ(rE(Uk 2(Uk+1 Uk)))Uk+1 Uk;Uk+1 Uki;forsome1and2in(0;1).SinceEisconvex,thenJ(rE)ispositivedeniteanditseigenvaluesarepositive.ByboundingtheeigenvaluesofJ(rE)by^0andaddingtheaboveexpressionswegethrE(Uk+1) rE(Uk);Uk+1 Uki^kUk+1 Ukk2:Substitutingthisin(2.9),weobtainE(Uk+1) E(Uk) ^ jj+1 tkUk+1 Ukk2:Byapplyingcondition(2.7)(i.e.,(2.8))theresultfollowsforallt0.Hencethemethodisunconditionallygradientstable.Toprovetheuniquesolvabilityof(2.4)weconsiderthenonlinearequationsUk+1+trEc(Uk+1)=Rk;whichmustbesolvedateachstepforagivenRk.SinceEcisstrictlyconvex,1 2kUk+1k2+tEc(Uk+1) hUk+1;RkihasauniqueminimuminUk+1forallt,and(2.4)hasauniquesolutionforallt0.Theconsistencyandthelocaltruncationerrorof(2.4)canbeestablishedbysimilarTaylorexpansionsastheoneswedidabovetoprovetheunconditionalstabilityofthescheme.MorepreciselyitconsistsofexpandingUk+1andUkaround(k+1=2)t,andrEc(Uk+1)andrE(Uk)aroundUk+1=2.ThisnishestheproofofTheorem2.3. Inthefollowingweapplytheideaofconvexitysplittingtoourthreeinpaintingmodels(1.4),(1.5),and(1.6).Forthiswechangefromthediscretesettingtothecontinuoussetting,i.e.,consideringfunctionsuinasuitableHilbertspaceinsteadofvectorsuinRN.Althoughthersttwooftheseinpaintingapproaches,i.e.,Cahn-HilliardinpaintingandTV-H 1inpainting,arenotgivenbygradient\rows,weshowthattheresultingnumericalschemesarestillunconditionallystable(inthesenseofDenition2.2)andthereforesuitabletosolvethemaccuratelyandreasonablyfast.ForinpaintingwithLCIS(1.6)theresultsofEyrecanbedirectlyapplied,eveninthecontinuoussetting;cf.Remark2.1.Nevertheless,alsoforthiscase,weadditionallypresentarigorousanalysis,similartotheonedoneforCahn-HilliardandTV-H 1inpainting.3.Cahn-HilliardinpaintingInthissectionweshowtheapplicationofconvexitysplittingtoCahn-Hilliardinpainting(1.4).Recallthattheinpaintedversionu(x)off(x)isconstructedbyfollowingtheevolutionequationu= u+1 F0(u)+(f u) 424UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGtosteadystate.ThismodiedCahn-Hilliardequationisintroducedin[13]fortheinpaintingofbinaryimages.Thelatter,mainlynumericalpaper,wasfollowedbyaverycarefulanalysisof(1.4)in[14].Tostartwith,theauthorsproveglobalex-istenceofauniqueweaksolutionoftheevolutionEquation(1.4).MorepreciselythesolutionuisproventobeanelementinC([0;T];L2(\n))\L2(0;T;V),whereV=2H2(\n)j@=@=0on@\n ,andistheoutwardpointingnormalon@\n.Underadditionalconditionsonthegivenimagef,theyalsoderivesomeveryin-terestingresultsconcerningthecontinuationofthegradientoftheimageintotheinpaintingdomain.Infact,in[14]theauthorsprovethatinthelimit0!1astationarysolutionof(1.4)solvesu 1 F0(u)=0;inD;u=f;on@D;ru=rf;on@D;(3.1)forfregularenough(f2C2).Theexistenceofastationarysolutionof(1.4)isassuredin[19].Additionally,in[14]theauthorspresentnumericalexampleswhichshowthattheconnectivityprincipleisfullled,andcomputeabifurcationdiagramforstationarysolutionsof(1.4).Thissupportstheclaimthatfourth-ordermethodsaresuperiortosecond-ordermethodswithrespecttoasmoothcontinuationoftheimagecontentsintothemissingdomain.Theideatoapplyconvexitysplittinginordertosolve(1.4)numericallywasbornin[13].Thenumericalresultspresentedthereillustratetheusefulnessofthisscheme.Althoughtheauthorsdonotanalyzetheschemerigorously,basedontheirnumericalresultstheyconjectureunconditionalstability.Inthefollowingweshallpresentthisnumericalschemeandderivesomeadditionalpropertiesbasedonarigorousanalysisofthelatter.TheoriginalCahn-Hilliardequationisagradient\rowinH 1fortheenergyE1(u)=Z\n 2jruj2+1 F(u)dx;whilethettingtermin(1.4)canbederivedfromagradient\rowinL2fortheenergyE2(u)=1 2Z\n(f u)2dx:However,notethatEquation(1.4)asawholeisnolongeragradientsystem.Hence,forthediscretizationintime,weapplytheconvexitysplittingdiscussedinSection2tobothfunctionalsE1andE2separately.Namely,wesplitE1asE1=E1c E1,withE1c(u)=Z\n 2jruj2+C1 2juj2dx;E1(u)=Z\n 1 F(u)+C1 2juj2dx:ApossiblesplittingforE2isE2=E2c E2withE2c(u)=1 2Z\nC2juj2dx;E2(u)=1 2Z\n (f u)2+C2juj2dx:TomakesurethatE1c;E1andE2c;E2arestrictlyconvex,theconstantsC1andC2havetobechosensuchthatC11 ,C20;see[14]. C.B.SCHONLIEBANDA.BERTOZZI425Thentheresultingdiscretetime-steppingschemeforaninitialconditionU0=u0isgivenbyUk+1 Uk t= rH 1(E1c(Uk+1) E1(Uk)) rL2(E2c(Uk+1) E2(Uk));whererH 1andrL2representgradientdescentwithrespecttotheH 1innerprod-uctandtheL2innerproductrespectively.ThistranslatestoanumericalschemeoftheformUk+1 Uk t+Uk+1 C1Uk+1+C2Uk+1=1 F0(Uk) C1Uk+(f Uk)+C2Uk;in\n:(3.2)WeenforceNeumannboundaryconditionson@\n,i.e.,rUk+1~n=rUk+1~n=0;on@\n;(3.3)where~nistheoutwardpointingnormalon@\n,andcomputeUk+1in(3.2)inthespectraldomainusingthediscretecosinetransform(DCT).TheideatousespectralmethodsforequationsinvolvingLaplacianoperatorsisclassicalandisbasedonthefactthattheLaplacematrixisdiagonalizedinthespectraldomain.Hence,solvingtheseequationsinthespectraldomaincanbedonemuchfastersincematrixmul-tiplicationisreplacedbyscalarmultiplication(multiplyingwiththeelementsinthemaindiagonal).Sinceadditionallytherealsoexistfastnumericalmethodstocom-putethediscreteFourier/Cosinetransform(suchasthefastFouriertransform(FFT))thismethodhasanoverallcomputationaladvantage.Let^UbetheDCTofUwitheigenvalues.ThenEquation(3.2)in^Ureads^Uk+1(i;j)=(1 C1t(1 x2+1 y2j)+C2t)^Uk(i;j)+ \F0(Uk)(i;j)+t\(f Uk) 1+C2t+t(1 x2+1 y2j)2 C1t(1 x2+1 y2j):3.1.RigorousEstimatesfortheScheme.FromTheorem2.3weknowthat(atleastinthespatiallydiscreteframework)theconvexitysplittingscheme(2.2)-(2.4)isunconditionallystable,i.e.,separatenumericalschemesforthegradient\rowsoftheenergiesE1(u)andE2(u)arenon{increasingforallt0.Butthisdoesnotguaranteethatthenumericalscheme(3.2)isunconditionallystable,sinceitcombinesthe\rowsoftwoenergies.Inthissectionweshallanalyzetheschemeinmoredetailandderivesomerigorousestimatesforitssolutions.Inparticularweshowthatthescheme(3.2)isunconditionallystableinthesenseofDenition2.2.Ourresultsaresummarizedinthefollowingtheorem.Theorem3.1.Letubetheexactsolutionof(1.4)anduk=u(kt)theexactsolutionattimekt,foratimestept0andk2N.LetUkbethekthiterateof(3.2)withconstantsC11=,C20.Thenthefollowingstatementsaretrue:(i)Undertheassumptionthatkuk 1andkruk2arebounded,thenumericalscheme(3.2)isconsistentwiththecontinuousEquation(1.4)andoforderoneintime. 426UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGUndertheadditionalassumptionthatF00(Uk 1)K(3.4)foranonnegativeconstantK,wefurtherhave(ii)ThesolutionsequenceUkisboundedonanitetimeinterval[0;T],forallt0.InparticularforktT,T0xed,wehaveforeveryt0krUkk22+tK1kUkk22eK2TkrU0k22+tK1kU0k22+tTC(\n;D;0;f);(3.5)forsuitableconstantsK1andK2,andconstantCdependingon\n;D;0;fonly.(iii)Thediscretizationerrorek,givenbyek=uk Uk,convergestozeroast0.Inparticular,wehaveforktT,T0xed,thatkrekk22+tC1 ~Ckekk22T ~CeK1TC(t)2;(3.6)forsuitableconstantsC;~C;K1.Remark3.1.NotethatourassumptionsfortheconsistencyofthenumericalschemeonlyholdifthetimederivativeofthesolutionofthecontinuousEquation(1.4)isuniformlybounded.Thisistrueforsmoothandboundedsolutionsoftheequation.Further,sinceweareinterestedinboundedsolutionsUkofthediscreteEquation(3.2),itisnaturaltoassume(3.4),i.e.,thatthenonlinearityF00intheprevioustimestep(k 1)tisbounded.AlsonotethattheconstantKin(3.4)canbechosenarbitrarilylarge.TheproofofTheorem3.1isorganizedinPropositions3.2-3.4.Proposition3.2(Consistency(i)).UnderthesameassumptionsasinTheorem3.1,andinparticularundertheassumptionthatkuk 1andkruk2arebounded,thenumericalscheme(3.2)isconsistentwiththecontinuousEquation(1.4)withkkk 1=O(t)ast0,wherekisthelocaltruncationerrorasdenedinEquation(2.6)above.Proof.Letkbethelocaltruncationerrordenedasin(2.6).Thenk=1k+2k;with1k=uk+1 uk t u(kt)2k=(uk+1 uk) C1(uk+1 uk)+C2(uk+1 uk)=t2uk+1 uk t C1tuk+1 uk t+C2tuk+1 uk t;i.e.,k=uk+1 uk t+2uk+1 1 F0(uk) (f uk) C1(uk+1 uk)+C2(uk+1 uk):(3.7) C.B.SCHONLIEBANDA.BERTOZZI427UsingstandardTaylorseriesargumentsandassumingthatkuk 1andkruk2areboundedwededucethattheglobaltruncationerrorisgivenby=maxkkkk 1=O(t)ast0:(3.8) Proposition3.3(Unconditionalstability(ii)).UnderthesameassumptionsasinTheorem3.1andinparticularassumingthat(3.4)holds,thesolutionsequenceUkfullls(3.5).Thisgivesboundednessofthesolutionsequenceon[0;T].Proof.WeconsiderourdiscretemodelUk+1 Uk t+Uk+1 C1Uk+1+C2Uk+1=1 F0(Uk) C1Uk+(f Uk)+C2Uk;multiplytheequationwith Uk+1andintegrateover\n.Weobtain1 tkrUk+1k22 hrUk;rUk+1i2+krUk+1k22+C1kUk+1k22+C2krUk+1k22=1 hF00(Uk)rUk;rUk+1i2+C1hUk;Uk+1i2+hr(f Uk);rUk+1i2+C2hrUk;rUk+1i2:UsingYoung'sinequalityweobtain1 2tkrUk+1k22 krUkk22+krUk+1k22+C1kUk+1k22+C2krUk+1k221 2kF00(Uk)rUkk22+ 2krUk+1k22+C1 2kUkk22+C1 2kUk+1k22+C2 2krUkk22+C2 2krUk+1k22+1 2kr(f Uk)k22+1 2krUk+1k22:Usingtheestimatekr(f Uk)k22220krUkk22+C(\n;D;0;f)andreorderingtheterms,weobtain1 2t+C2 2 1 2krUk+1k22+C1 2kUk+1k22+ 2krUk+1k221 2t+C2 2+20krUkk22+1 2kF00(Uk)rUkk22+C1 2kUkk22+C(\n;D;0;f):Bychoosing=22,thethirdtermontheleftsideoftheinequalityiszero.BecauseofAssumption(3.4)weobtainthefollowingboundontherightsideoftheinequalitykF00(Uk)rUkk22K2krUkk22;andwehave1 2t+C2 2 1 2krUk+1k22+C1 2kUk+1k221 2t+C2 2+20+K2 43krUkk22+C1 2kUkk22+C(\n;D;0;f): 428UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGNowwemultiplytheaboveinequalityby2tanddene~C=1+t(C2 1);~~C=1+tC2+220+K2 23:SinceC2ischosengreaterthan01,therstcoecient~Cispositiveandwecandividetheinequalitybyit.WeobtainkrUk+1k22+tC1 ~CkUk+1k22~~C ~CkrUkk22+tC1 ~CkUkk22+tC(\n;D;0;f);whereweupdatedtheconstantC(\n;D;0;f)byC(\n;D;0;f)=~C.Since~~C ~C1,wecanmultiplythesecondtermontherightsideoftheinequalitybythisquotienttoobtainkrUk+1k22+tC1 ~CkUk+1k22~~C ~CkrUkk22+tC1 ~CkUkk22+tC(\n;D;0;f):WededucebyinductionthatkrUkk22+tC1 ~CkUkk22 ~~C ~C!kkrU0k22+tC1 ~CkU0k22+tk 1X=0 ~~C ~C!C(\n;D;0;f)=(1+K2t)k (1+K1t)kkrU0k22+tC1 ~CkU0k22+tk 1X=0(1+K2t) (1+K1t)C(\n;D;0;f):ForktTwehavekrUkk22+tC1 ~CkUkk22e(K2 K1)TkrU0k22+tC1 ~CkU0k22+tTe(K2 K1)TC(\n;D;0;f)=e(K2 K1)TkrU0k22+tC1 ~CkU0k22+tTC(\n;D;0;f);whichgivesboundednessofthesolutionsequenceon[0;T]foranyT0,assumingthat(3.4)holds. Theconvergenceofthediscretesolutiontothecontinuousoneasthetimestept0isveriedinthefollowingproposition.Proposition3.4(Convergence(iii)).UnderthesameassumptionsasinTheorem3.1andinparticularunderAssumption(3.4)thediscretizationerrorekfullls(3.6). C.B.SCHONLIEBANDA.BERTOZZI429InordertoproveProposition3.4weneedthefollowingauxiliarylemma.Lemma3.2.TheerrorekbetweentheexactandapproximatesolutiondenedasinTheorem3.1fulllsZ\nekdx=O((t)2):Proof.[ProofofLemma3.2]Becauseofthedelitytermin(1.4)and(3.2),solutionsoftheseequationsarenotmasspreserving,i.e.,R\nekdoesnotingeneralvanish.Infactwehave,forasolutionukof(1.4),d dtZ\nuk= Z\n2uk+1 Z\nF0(uk)+Z\n(f uk)= Z@\nruk~n+1 Z@\nrF0(uk)~n+Z\n(f uk);wherewehaveusedGaussdivergencetheoremtoobtaintheboundaryintegrals.AssumingzeroNeumannboundaryconditionsasin(3.3)thetwoboundaryintegralsvanish,andhenced dtZ\nuk=Z\n(f uk):Inparticulard dtZDuk=0:(3.9)Asimilarcomputationforthediscretesolutionof(3.2)showsthat1 t+C2Z\n(Uk+1 Uk)=Z\n(f Uk);andinparticular1 t+C2ZD(Uk+1 Uk)=0:(3.10)Next,letusfollowthelinesoftheconsistencyproofin(3.7).Thenthediscretizationerroreksatisesek+1 ek t+2ek+1 C1ek+1+C2ek+1=1 t(uk+1 uk) 1 t(Uk+1 Uk)+2uk+1 2Uk+1 C1uk+1+C1Uk+1+C2uk+1 C2Uk+1= 1 F0(Uk) C1Uk+(f Uk)+C2Uk+1 F0(uk)+(f uk) C1uk+C2uk+k= 1 (F0(Uk) F0(uk)) C1(Uk uk)+C2(Uk uk) (Uk uk)+k: 430UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGAsbefore,integratingover\n,applyingGaussdivergencetheoremandthezeroNeu-mannboundaryconditionsforukandUk,weget1 t+C2Z\n(ek+1 ek)+Z\nek=Zk;(3.11)whereZ\nk=1 t+C2Z\n(uk+1 uk) Z\n(uk)=1 t+C2Z\n(uk+t(uk)+O((t)2) uk) Z\n(uk)=O(t):Now,toproveourclaimweapplyinductiononk.First,assumingthatu0=U0in\n,wehavethatZ\ne0=0;andhence1 t+C2Z\ne1=O(t):(3.12)Assumingthatassertion(3.12)holdsforallindiceskandusing(3.9)and(3.10)wehave,for(3.11),1 t+C2Z\n(ek+1 ek)+Z\nek=Zk1 t+C2Z\nek+1 O(t)+01 t+C2 1O(t)=O(t)1 t+C2Z\nek+1=O(t);andhence(1+C2t)Z\nek=O((t)2)forallk0. WecontinuewiththeproofofProposition3.4.Proof.[ProofofProposition3.4]IntheproofofLemma3.2wehaveusedtheconsistencyresult(3.7)toshowthatthediscretizationerroreksatisesek+1 ek t+2ek+1 C1ek+1+C2ek+1= 1 (F0(Uk) F0(uk)) C1(Uk uk)+C2(Uk uk) (Uk uk)+k: C.B.SCHONLIEBANDA.BERTOZZI431Multiplicationwith ek+1leadsto1 thr(ek+1 ek);rek+1i2+krek+1k22+C1kek+1k22+C2krek+1k22=1 h(F0(Uk) F0(uk));ek+1i2 C1h(Uk uk);ek+1i2+hr(Uk uk);rek+1i2 C2hr(Uk uk);rek+1i2+\nr 1k;rek+12:Further,because1 tkrek+1k22 hrek;rek+1i21 2t(krek+1k22 krekk22);weobtain1 2t(krek+1k22 krekk22)+krek+1k22+C1kek+1k22+C2krek+1k221 h(F0(Uk) F0(uk));ek+1i2+C1hek;ek+1i2 hrek;rek+1i2+C2hrek;rek+1i2+\nr 1k;rek+12:ApplyingYoung'sinequalityleadsto1 2t(krek+1k22 krekk22)+krek+1k22+C1kek+1k22+C2krek+1k22 1 h(F00(Uk)rUk F00(uk)ruk);rek+1i2+C1 21kekk22+C11 2kek+1k22+20 23krekk22+3 2krek+1k22+C2 22krekk22+C22 2krek+1k22+1 24kkk2 1+4 2krek+1k22:Letusconsidertheremaininginnerproductinthelastinequality: 1 h(F00(Uk)rUk F00(uk)ruk);rek+1i2=1 hF00(Uk)rek;rek+1i2+1 h(F00(uk) F00(Uk))ruk;rek+1i21 25kF00(Uk)jrekjk22+1 26k(F00(uk) F00(Uk))jrukjk22+5 2+6 2krek+1k22:Nextweassumethat(3.4)holdsandthatrukisuniformlyboundedon[0;T]|inparticular,that9~~K0suchthatkrukk2~~KforallktT:(3.13)ThelatterassumptionwillbeproveninLemma3.3justaftertheendofthisproof.Moreover,sinceF00islocallyLipschitzcontinuousweobtain 1 h(F00(Uk)rUk F00(uk)ruk);rek+1i2C 25krekk22+C 26kekk22+5 2+6 2krek+1k22; 432UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGwherewehavesetCtobeauniversalconstantforallbounds.Further,usingLemma3.2andkekk22=kek O(t)2+O(t)2k222kek O(t)2k22+2kO(t)2k22;wecanapplythePoincareinequalitytotheL2normofek.Insumweget1 2t+C21 2 2 3 2krek+1k22+C11 1 2kek+1k22+ 4 2 5+6 2krek+1k221 2t+20 23+C2 22+C 25+C 6krekk22+C1 21kekk22+1 24kkk2 1+C 6kO(t)2k22:Nextwechoose1=1andmultiplytheinequalitywith2t:(1+t(C2(2 2) 3))krek+1k22+tC1kek+1k22+t2 4 5+6 krek+1k221+t20 3+C2 2+C 5+2C 6krekk22+tC1kekk22+t 4kkk2 1+t2C 6kO(t)2k22:Let~C=1+t(C2(2 2) 3);~~C=1+t20 3+C2 2+C 5+2C 6:Now,bychoosingallssuchthatthecoecientsofalltermsintheinequalityarenonnegativeandthequotient~~C=~C1,andbyestimatingthelasttermontheleftsidefrombelowbyzero,wegetkrek+1k22+tC1 ~Ckek+1k22~~C ~Ckrekk22+tC1 ~Ckekk22+t ~C1 4kkk2 1+2C 6kO(t)2k22;andbecause~~C=~C1wefurtherhavekrek+1k22+tC1 ~Ckek+1k22~~C ~Ckrekk22+tC1 ~Ckekk22+t ~C1 4kkk2 1+2C 6kO(t)2k22: C.B.SCHONLIEBANDA.BERTOZZI433Byinductiononkweobtainkrek+1k22+tC1 ~Ckek+1k22 ~~C ~C!k+1kre0k22+tC1 ~Cke0k22+t ~CkX=0 ~~C ~C!1 4maxkfkk2 1g+2C 6kO(t)k22=t ~CkX=0(1+K1t)1 4maxkfkk2 1g+2C 6kO(t)2k22t ~CkeK1k1 4maxkfkk2 1g+2C 6kO(t)2k22;wherewehaveusedthefactthate0=0and1~~C ~C=1+K1t.Hence,byusingtheconsistencyresult(3.8)weconclude,forktT,thatkrekk22+tC1 ~Ckekk22T ~CeK1TC(t)2: From[13,14]weknowthatthesolutionuktothecontinuousequationgloballyexistsandisuniformlyboundedinL2(\n).Nextweshowthatassumption(3.13)holds.Lemma3.3.Letukbetheexactsolutionof(1.4)attimet=ktandletT0.ThenthereexistsaconstantC0suchthatkrukk2CforallktT.Proof.LetK(u)= u+1 F0(u).WemultiplythecontinuousevolutionEqua-tion(1.4)withK(u)andobtainhu;K(u)i2=hK(u);K(u)i2+h(f u);K(u)i2:LetusfurtherdeneE(u):= 2Z\njruj2dx+1 Z\nF(u)dx:Thenwehavehu;K(u)i2=u; u+1 F0(u)2=hru;rui2+u;1 F0(u)2=d dtE(u);sinceusatisesNeumannboundaryconditions.Thereforewegetd dtE(u)= Z\njrK(u)j2dx+h(f u); ui2+(f u);1 F0(u)2:(3.14) 434UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGSinceF(u)isboundedfrombelow,weonlyhavetoshowthatE(u)isuniformlyboundedon[0;T],andweautomaticallyhavethatjrujisuniformlyboundedon[0;T].Westartwiththelastterm,andrecallthefollowingboundsonF0(u)(cf.[67]):ThereexistpositiveconstantsC1;C2suchthatF0(s)sC1s2 C2;8s2Rand,forevery0,thereexistsaconstantC3suchthatjF0(s)jC1s2+C3();8s2R:Usingthelasttwoestimatesweobtainthefollowing:(fu);1 F0(u)20 Z\nnDF0(u)fdx0 Z\nnDF0(u)udx0 Z\nnDjF0(u)jdxkfkL1(\n)0C1 Z\nnDu2dx0C2j\nnDj 0C(f;\n) C1 Z\nnDu2dxC3()j\nnDj !0C1 Z\nnDu2dx0C2j\nnDj 0C1 (1C(f;\n))Z\nnDu2dxC(0;;;\n;D;f);wherewechoose1=C(f;\n).Thereforeintegrating(3.14)overthetimeinterval[0;T]resultsinZT0d dtE(u(t))dtZT0 Z\njrK(u)j2dxdt+ZT0h(f u); ui2dt 0C1 (1 C(f;\n))ZT0Z\nnDu2dxdt+TC(0;;;\n;D;f):Nextweconsiderthesecondtermontherightsideofthelastinequality.FromTheorem4.1in[13]weknowthatasolutionuof(1.4)isanelementinL2(0;T;H2(\n))forallT0.Henceu2L2(0;T;L2(\n))andthesecondtermisboundedbyaconstantdependingonT.Consequently,foreach0tT,wegetE(u(t))E(u(0))+C(T)+TC(0;;;\n;D;f) ZT0"Z\njrK(u)j2dx+0C1 (1 C(f;\n))Z\nnDu2dx#dt;andwiththis,foraxedT0,thatjrujisuniformlyboundedin[0;T]. 3.2.Numericalresults.Inourcomputationstheoptimaltturnedouttobet=1or10(dependingalsoonthesizeofand0).NumericalresultsoftheaboveschemearepresentedinFigures3.1,3.2and3.3.Inalloftheexampleswefollowtheprocedureof[13],i.e.,theinpaintedimageiscomputedinatwostepprocess.IntherststepCahn-Hilliardinpaintingissolvedwitharatherlargevalueof,e.g.,=0:1,untilthenumericalschemeisclosetosteadystate.Inthisstepthelevellinesarecontinuedintothemissingdomain.Inasecondstep,theresultofthe C.B.SCHONLIEBANDA.BERTOZZI435 Fig.3.1.BinaryimagewithunknowncenterandthesolutionofCahn-Hilliardinpaintingwith0=105andswitchingvalue:u(600)with=0:1,u(1000)with=0:01 Fig.3.2.Textremovalfromabinaryimage:thesolutionofCahn-Hilliardinpaintingwith0=109andswitchingvalue:u(200)with=0:8,u(500)with=0:01rststepisputasaninitialconditionintotheschemeforasmall,e.g.,=0:01,inordertosharpenthecontoursoftheimagecontents.Thereasonforthistwostepprocedureistwofold.Firstofallin[14]theauthorsgivenumericalevidencethatthesteadystateofthemodiedCahn-HilliardEquation(1.4)isnotunique,i.e.,itisdependentontheinitialconditioninsideoftheinpaintingdomain.Asaconsequence,computingtheinpaintedimagebytheapplicationofCahn-Hilliardinpaintingwithasmallonly,mightnotextendthelevellinesintothemissingdomainasdesired.Seealso[14]forabifurcationdiagrambasedonthenumericalcomputationsoftheauthors.ThesecondreasonforsolvingCahn-Hilliardinpaintingintwostepsisthatitiscomputationallylessexpensive.Solvingtheabovetime-marchingschemefor,e.g.,=0:1isfasterthansolvingitfor=0:01.ThisisbecauseofthedampingintroducedbyC1,i.e.,,intothescheme;cf.(3.2).Allnumericalexamplespresentedherehavebeencomputedinordersof10secondsona1.86GHzprocessorwith1GBRAM.ForafurtherdiscussiononcomputationaltimesfortheconvexitysplittingmethodappliedtoCahn-Hilliardinpaintingwereferto[13].OnepossiblegeneralizationofCahn-Hilliardinpaintingforgrayscaleimagesistosplitthegrayscaleimagebit-wiseintochannelsu(x)KXk=1uk(x)2 (k 1);whereK0.TheCahn-Hilliardinpaintingapproachisthenappliedtoeachbinarychannelukseparately,compareFigure3.5.Attheendoftheinpaintingprocessthechannelsareassembledagainandtheresultistheinpaintedgrayvalueimageinlowergrayvalueresolution,compareFigure3.4.InFigure3.6theapplicationofbitwiseCahn-Hilliardinpaintingfortherestorationofsatelliteimagesofroadsis 436UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTING Fig.3.3.VandalizedbinaryimageandthesolutionofCahn-Hilliardinpaintingwith0=109andswitchingvalue:u(800)with=0:8,u(1600)with=0:01demonstrated.Onecanimaginethattheblackdotsintherstpicturerepresenttreesthatcoverpartsoftheroad.Theideaofbitwisebinaryinpaintingisproposedin[30]fortheinpaintingwithwaveletsbasedontheAllen-Cahnenergy. 20 40 60 80 100 120 20 40 60 80 100 120 140 160 Inpainted image 20 40 60 80 100 120 20 40 60 80 100 120 140 160 Fig.3.4.Cahn-HilliardbitwiseinpaintingwithK=8binarychannels(0=108,with=0:1untilt=800and=0:01untilt=1200)4.TV-H 1inpaintingInthissectionwediscussconvexitysplittingforTV-H 1inpainting(1.5).ToavoidnumericalandtheoreticaldicultiesweapproximateanelementpinthesubdierentialofthetotalvariationfunctionalTV(u)byasmoothedversionofr(ru=jruj),thesquarerootregularizationforinstance.Withthelatterregu-larizationthesmoothedversionof(1.5)readsasu= r ru p jruj2+2!+(f u);(4.1)with01.Incontrasttoitssecond-orderanalogue,thewell-posednessof(1.5)stronglydependsonthesmoothingusedforr(ru=jruj).Infacttherearesmoothingfunctionsforwhich(1.5)producessingularitiesinnitetime.Thisiscausedbythelackofmaximumprincipleswhichinthesecond-ordercaseguaranteethewell-posednessforallsmoothmonotoneregularizations.In[16]theauthorsconsider(1.5)with=0inallof\n,i.e.,thefourth-orderanaloguetoTV-L2denoising,whichwasoriginallyintroducedin[58].Theyproveglobalwell-posedness,inonespace C.B.SCHONLIEBANDA.BERTOZZI437 initial condition2 20 40 60 80 100 120 20 40 60 80 100 120 140 160 initial condition3 20 40 60 80 100 120 20 40 60 80 100 120 140 160 initial condition5 20 40 60 80 100 120 20 40 60 80 100 120 140 160 1200 iterations,2channel 20 40 60 80 100 120 20 40 60 80 100 120 140 160 1200 iterations,3channel 20 40 60 80 100 120 20 40 60 80 100 120 140 160 1200 iterations,5channel 20 40 60 80 100 120 20 40 60 80 100 120 140 160 Fig.3.5.Thegivenimage(rstrow)andtheCahn-Hilliardinpaintingresult(secondrow)forthechannels2;3and5. Fig.3.6.BitwiseCahn-HilliardinpaintingwithK=8binarychannelsappliedtoroadrestorationdimensionandforsmoothinitialdata,forthearctanregularization2 arctan(ux=)x;(4.2)where01.Forthesquarerootsmoothing ux p juxj2+2!x(4.3)theyconjecture,supportedbyempiricalevidence,thatsingularitiesoccurininnitetime,notnitetime.Thebehaviorofthefourth-orderPDEinonedimensionisalsorelevantfortwo-dimensionalimagessincealotofstructureinvolvesedgeswhichareone-dimensionalobjects.Intwodimensionssimilarresultsaremuchmorediculttoobtain,sinceenergyestimatesandtheSobolevlemmainvolvedinitsproofmightnot 438UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGholdinhigherdimensionsanymore.Wealsonotethatin[19]theauthorsprovetheexistenceofaweakstationarysolutionfor(1.5)intwospacedimensions.Inthefollowingwepresenttheconvexitysplittingmethodappliedto(1.5)forboththesquarerootandthearctanregularization.SimilarlytotheconvexitysplittingforCahn-Hilliardinpainting,weproposethefollowingsplittingfortheTV-H 1inpaintingequation.Theregularizingtermin(1.5)canbemodeledbyagradient\rowinH 1oftheenergyE1(u)=Z\njrujdx;wherejrujisreplacedbyitsregularizedversion,e.g., jruj2+2,0.WesplitE1asE1c E1,withE1c(u)=Z\nC1 2jruj2dx;andE1(u)=Z\n jruj+C1 2jruj2dx:ThettingtermissplitintoE2=E2c E2,analogoustoCahn-Hilliardinpainting.Theresultingtime-steppingschemeisgivenbyUk+1 Uk t+C1Uk+1+C2Uk+1=C1Uk rrUk jrUkj+C2Uk+(f Uk):(4.4)WeassumethatUk+1satiseszeroNeumannboundaryconditionsandusetheDCTtosolve(4.4).TheconstantsC1andC2havetobechosensuchthatE1c;E1;E2c;E2areallstrictlyconvex.Inthefollowingwedemonstratehowtocomputetheappropriateconstants.LetusconsiderC1rst.ThefunctionalE1cisstrictlyconvexforallC10.ThechoiceofC1fortheconvexityofE1dependsontheregularizationofthetotalvariationweareusing.Weusethesquareregularization(4.3),i.e.,insteadofjrujwehaveZG(jruj)dx;withG(s)= s2+2:Settingy=jrujwehavetochooseC1suchthatC1 2y2 G(y)isconvex.TheconvexityconditionforthesecondderivativegivesusthatC1G00(y)()C12 (2+y2)3=2()C11 issucientas2 (2+y2)3=2hasitsmaximumvalueaty=0.Intheonedimensionalcase,wewouldliketocomparethiswiththearctanregularization(4.2),i.e.,replacingx jxjby2 arctan(x )asproposedin[16].HeretheconvexityconditionforthesecondderivativereadsC1d ds2 arctans 0:Thesignresultsfromtheabsentabsolutevalueintheregularizationdenition.WeobtainC12 1 (1+s2=2)0: C.B.SCHONLIEBANDA.BERTOZZI439TheinequalitywithaplussigninsteadofistrueforallconstantsC10.IntheothercaseweobtainC12 2+s2;whichisfullledforalls2RifC12 .Notethatthisconditionisalmostthesameasinthecaseofthesquareregularization.NowweconsiderE2=E2c E2.ThefunctionalE2cisstrictlyconvexifC20.FortheconvexityofE2werewriteE2(u)=1 2Z\n (f u)2+C2juj2dx=ZDC2 2juj2dx+Z\nnD 0 2(f u)2+C2 2juj2dx=ZDC2 2juj2dx+Z\nnDC2 2 0 2juj2+0fu 0 2jfj2:ThisisconvexforC20,e.g.,withC2=0+1wecanwriteE2(u)=ZDC2 2juj2dx+Z\nnD1 2u+0f2 20+0 2jfj2dx:4.1.Rigorousestimatesforthescheme.AsinSection3.1forCahn-Hilliardinpainting,weproceedwithamoredetailedanalysisof(4.4).Throughoutthissectionweconsiderthesquare-rootregularizationofthetotalvariationbothinournumericalschemeandinthecontinuousevolutionEquation(1.5).Notethatsimilarresultsaretrueforothermonotoneregularizerssuchasthearctansmoothing.Ourresultsaresummarizedinthefollowingtheorem.Theorem4.1.Letubetheexactsolutionof(4.1)anduk=u(kt)betheexactsolutionattimektforatimestept0andk2N.LetUkbethekthiterateof(4.4)withconstantsC11=,C20.Thenthefollowingstatementsaretrue:(i)Undertheassumptionthatkuk 1andkruk2arebounded,thenumericalscheme(4.4)isconsistentwiththecontinuousEquation(1.5)andoforderoneintime.(ii)ThesolutionsequenceUkisboundedonanitetimeinterval[0;T],forallt0.Inparticular,forktT,T0xed,wehaveforeveryt0,krUkk22+tK1krUkk22eK2T krU0k22+tK1krU0k22+tTC(\n;D;0;f)(4.5)forsuitableconstantsK1,K2,andaconstantC,whichdependson\n;D;0;fonly.(iii)Letek=uk Uk.ForsmoothsolutionsukandUk,theerrorekconvergestozeroast0.Inparticular,forktT,T0xed,wehavekrekk22+tM1krekk22T M2eM3T(t)2(4.6)forsuitablepositiveconstantsM1;M2andM3. 440UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGRemark4.1.FortheconvergenceresultinTheorem4.1(iii)weassumethatsmoothsolutionstoboththecontinuousintimeproblemandthediscreteintimeapproximationexist.Thevalidityofthisassumptionisnotknowningeneral.Notehoweverthattheglobalregularityresultsareknownin1Dforthearctansmooth-ing[16].Moreover,ournumericalresultsshownoindicationofsingularitiesin2D.Therefore,itisnotunreasonabletoanalyzetheconvergenceundertheseassumptions.TheproofofTheorem4.1issplitintothethreeseparatePropositions4.2-4.4.Proposition4.2(Consistency(i)).UnderthesameassumptionsasinThe-orem4.1andinparticularundertheassumptionthatkuk 1andkruk2arebounded,thenumericalscheme(4.4)isconsistentwiththecontinuousEquation(4.1)withkkk 1=O(t)ast0,wherekisthelocaltruncationerror.Proof.Thelocaltruncationerrorisdenedoveratimestepassatisfyingk=1k+2k;where1k=uk+1 uk t u(kt);2k=C12(uk+1 uk)+C2(uk+1 uk);i.e.,k=uk+1 uk t+ r ruk p jrukj2+2!! (f uk)+C12(uk+1 uk)+C2(uk+1 uk):(4.7)UsingstandardTaylorseriesargumentsandassumingthatkuk 1,kruk2andkuk2areboundedwededucethatkkk 1=O(t)fort0:(4.8) Proposition4.3(Unconditionalstability(ii)).Underthesameassump-tionsasinTheorem4.1thesolutionsequenceUkfullls(4.5).Thisgivesboundednessofthesolutionsequenceon[0;T].Proof.Ifwemultiply(4.4)with Uk+1andintegrateover\nweobtain1 tkrUk+1k22 hrUk;rUk+1i2+C2krUk+1k22+C1krUk+1k22=*r0@rUk q jrUkj2+21A;Uk+1+2+C1hrUk;rUk+1i2+hr((f Uk));rUk+1i2+C2hrUk;rUk+1i2:ApplyingYoung'sinequalitytotheinnerproductsontherightandestimatingkr(f Uk)k22220krUkk22+C(\n;D;0;f) C.B.SCHONLIEBANDA.BERTOZZI441resultsin1 2tkrUk+1k22 krUkk22+C2krUk+1k22+C1krUk+1k22*r0@rUk q jrUkj2+21A;Uk+1+2+C1 1krUkk22+C11krUk+1k22+220 2krUkk22+2krUk+1k22+C2 3krUkk22+C23krUk+1k22+C(\n;D;0;f):Now,thersttermontherightsideoftheinequalitycanbeestimatedasfollows*r0@rUk q jrUkj2+21A;Uk+1+2= *rr0@rUk q jrUkj2+21A;rUk+1+21 4\r\r\r\r\r\rrr0@rUk q jrUkj2+21A\r\r\r\r\r\r22+4krUk+1k22:ApplyingPoincare'sandCauchy'sinequalitytothersttermleadsto\r\r\r\r\r\rrr0@rUk q jrUkj2+21A\r\r\r\r\r\r22C (krUkk22+kUkk22+krUkk22):InterpolatingtheL2normofubytheL2normsofruandru,weobtain1 2t+C2(1 3) 2krUk+1k22+(C1(1 1) 4)krUk+1k221 2t+220 2+C2 3+C(1=;\n) 4krUkk22+C1 1+C(1=;\n) 4krUkk22+C(\n;D;0;f):For=1=2,i=1;:::;4weobtain1 2t+C2 1 2krUk+1k22+C1 1 2krUk+1k221 2t+420+2(C2+C)krUkk22+2(C1+C)krUkk22+C(\n;D;0;f):SinceC1andC2arechosensuchthatC11=1andC201,thecoecientsintheinequalityabovearepositive.TherestoftheproofissimilartotheproofofProposition3.3.Wemultiplytheinequalityby2tandsetCa=1+t(C2 1);C=C1 1;Cc=1+2t(420+2(C2+C));Cd=4(C1+C):WeobtainCakrUk+1k22+tCkrUk+1k22CckrUkk22+tCdkrUkk22+2tC(\n;D;0;f): 442UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGDividingbyCa(whichis0)wehavekrUk+1k22+tCb CakrUk+1k22Cc CakrUkk22+tCd CakrUkk222 CatC(\n;D;0;f):WerewritetherighthandsideoftheinequalitysuchthatkrUk+1k22+tC CakrUk+1k22CcCd Ca1 CdkrUkk22+t1 CckrUkk22+2 CatC(\n;D;0;f):SinceCd1wecanmultiplythersttermwithinthebracketsontherighthandsideoftheinequalitywithCdandwillonlygetsomethingwhichislargerorequal.Forthesamereasonwecanmultiplythesecondtermwithinthebracketswith1CcC Ca=(1+2t(420+2(C2+C)))(C1 1) 1+t(C2 1)andgetkrUk+1k22+tC CakrUk+1k22CcCd CakrUkk22+tC CakrUkk22+2 CatC(\n;D;0;f):ByinductionitfollowsthatkrUk+1k22+tC CakrUk+1k22CcCd CakkrU0k22+tC CakrU0k22+tk 1X=0CcCd Ca2 CaC(\n;D;0;f):ThereforeweobtainforktTkrUkk22+tCb CakrUkk22eKTkrU0k22+tCb CakrU0k22+tT2 CaC(\n;D;0;f): Finallyweshowthatthediscretesolutionconvergestothecontinuousoneasttendstozero.Proposition4.4(Convergence(iii)).UnderthesameassumptionsasinTheorem4.1(iii)theerrorekfullls(4.6).Proof.ByourdiscreteApproximation(4.4)andtheconsistencycomputation C.B.SCHONLIEBANDA.BERTOZZI443(4.7),wehaveforek=uk Ukek+1 ek t+C12ek+1+C2ek+1=1 t(uk+1 uk) 1 t(Uk+1 Uk)+C12uk+1 C12Uk+1+C2uk+1 C2Uk+1= 0@C12Uk 0@r0@rUk q jrUkj2+21A1A+(f Uk)+C2Uk1A 0@0@r0@ruk q jrukj2+21A1A (f uk) C12uk C2uk1A+k= " 0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A+C12(Uk uk)+C2(Uk uk) (Uk uk)#+k:Takingtheinnerproductwith ek+1,wehave1 thr(ek+1 ek);rek+1i2+C1krek+1k22+C2krek+1k22=* 0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A;ek+1+2+C1\n2(Uk uk);ek+12+hr(Uk uk);rek+1i2 C2hr(Uk uk);rek+1i2 \nr 1k;rek+12:UsingthesameargumentsasintheproofofProposition3.4weobtain1 2t(krek+1k22 krekk22)+C1krek+1k22+C2krek+1k22* 0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A;ek+1+2+C1 1krekk22+C11krek+1k22+20 3krekk22+3krek+1k22+C2 2krekk22+C22krek+1k22+1 4kkk2 1+4krek+1k22:(4.9)Weconsiderthersttermontherightsideoftheaboveinequalityindetail,* 0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A;ek+1+2=*r0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A;rek+1+2:(4.10) 444UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGWegetr0@ru q jruj2+21A=u q jruj2+2 u2xuxx+2uxuyuxy+u2yuyy (jruj2+2)3=2:Next,weapplythegradienttothisexpressionandobtainr0@r0@ru q jruj2+21A1A=ru q jruj2+2 u 2(jruj2+2)3=2rjruj2 r(u2xuxx+2uxuyuxy+u2yuyy) (jruj2+2)3=2+3(u2xuxx+2uxuyuxy+u2yuyy) 2(jruj2+2)5=2rjruj2;whererjruj2=0BB@2ruuxxuyx2ruuxyuyy1CCA=2uxuxx+uyuyxuxuxy+uyuyy;andr(u2xuxx+2uxuyuxy+u2yuyy)=2(uxuxx+uyuxy)rux+2(uxuxy+uyuyy)ruy+u2xruxx+2uxuyruxy+u2yruyy:Reorderingtheinvolvedtermswehaver0@r0@ru q jruj2+21A1A=H1(ru)ru+H2(ux;uy;uxx;uxy;uyy)rux+H3(ux;uy;uxx;uxy;uyy)ruy+H4(ux;uy)ruxx+H5(ux;uy)ruxy+H6(ux;uy)ruyy; C.B.SCHONLIEBANDA.BERTOZZI445whereH1(ru)=1 q jruj2+2;H2(ux;uy;uxx;uxy;uyy)= uux+2(uxuxx+uyuxy) (jruj2+2)3=2 3(u2xuxx+2uxuyuxy+u2yuyy)ux (jruj2+2)5=2!;H3(ux;uy;uxx;uxy;uyy)= uuy+2(uxuxy+uyuyy) (jruj2+2)3=2 3(u2xuxx+2uxuyuxy+u2yuyy)uy (jruj2+2)5=2!;H4(ux;uy)= u2x (jruj2+2)3=2;H5(ux;uy)= 2uxuy (jruj2+2)3=2;H6(ux;uy)= u2y (jruj2+2)3=2:Nowwearegoingtoinsertthisinto(4.10).Foreaseofnotationwesuppressthetimeindexkfornow,i.e.,wedeneU:=Uk,u:=ukande:=ek.Weobtain*r0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A;rek+1+2=hH1(rU)rU H1(ru)ru;rek+1i2+hH2(Ux;Uy;Uxx;Uxy;Uyy)rUx H2(ux;uy;uxx;uxy;uyy)rux;rek+1i2+hH3(Ux;Uy;Uxx;Uxy;Uyy)rUy H3(ux;uy;uxx;uxy;uyy)ruy;rek+1i2+hH4(Ux;Uy)rUxx H4(ux;uy)ruxx;rek+1i2+hH5(Ux;Uy)rUxy H5(ux;uy)ruxy;rek+1i2+hH6(Ux;Uy)rUyy H6(ux;uy)ruyy;rek+1i2 446UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTING1 2kH1(rU)(rU ru)k22+1 2kru(H1(ru) H1(rU))k22+1 2kH2(Ux;Uy;Uxx;Uxy;Uyy)(rUx rux)k22+1 2krux(H2(ux;uy;uxx;uxy;uyy) H2(Ux;Uy;Uxx;Uxy;Uyy))k22+1 2kH3(Ux;Uy;Uxx;Uxy;Uyy)(rUy ruy)k22+1 2kruy(H3(ux;uy;uxx;uxy;uyy) H3(Ux;Uy;Uxx;Uxy;Uyy))k22+1 2kH4(Ux;Uy)(rUxx ruxx)k22+1 2kruxx(H4(ux;uy) H4(Ux;Uy))k22+1 2kH5(Ux;Uy)(rUxy ruxy)k22+1 2kruxy(H5(ux;uy) H5(Ux;Uy))k22+1 2kH6(Ux;Uy)(rUyy ruyy)k22+1 2kruyy(H6(ux;uy) H6(Ux;Uy))k22+6krek+1k22;forasuitableconstant0.NextwewanttousethattheH'sareLipschitzcontin-uousin\n,withLipschitzconstantsL(1=),for-0.8;أ倀0,whichgrowasdecreases.Forsimplicity,weonlypresenttheprooffortherstpartofH2,i.e.,forH12(ux;uy;uxx;uxy;uyy)= uux+2(uxuxx+uyuxy) (jruj2+2)3=2=ux(3uxx+uyy)+2uyuxy (jruj2+2)3=2:Theothersfollowsimilarily.WehavekH12(ux;uy;uxx;uxy;uyy)H12(Ux;Uy;Uxx;Uxy;Uyy)k2\r\r\r\rux(3uxxuyy)+2uyuxy (jruj22)3=2Ux(3UxxUyy)+2UyUxy (jrUj22)3=2\r\r\r\r2\r\r\r\rux(3uxxuyy) (jruj22)3=2Ux(3UxxUyy) (jrUj22)3=2\r\r\r\r2+2\r\r\r\ruyuxy (jruj22)3=2UyUxy (jrUj22)3=2\r\r\r\r2\r\r\r\r(3uxxuyy)ux (jruj22)3=2Ux (jrUj22)3=2\r\r\r\r2\r\r\r\rUx (jrUj22)3=2(3exxeyy)\r\r\r\r2+2\r\r\r\ruxyuy (jruj22)3=2Uy (jrUj22)3=2\r\r\r\r2+2\r\r\r\rUy (jrUj22)3=2exy\r\r\r\r2:FromourassumptioninTheorem4.1(iii)wehaveacontinuousintimesmoothsolutionuonanitetimeinterval.Inparticularthisgivesusauniformboundforthesecondderivativesoftheexactsolutionu,i.e.,thereexistsaC0suchthatkuxxk1+kuxyk1+kuyyk1Conanitetimeinterval[0;T].Further,withthefactthatthefunctionx (x2+y2+2)3=2isuniformlyboundedfor0andforallx;y2RwehavekH12(ux;uy;uxx;uxy;uyy) H12(Ux;Uy;Uxx;Uxy;Uyy)k2C\r\r\r\r\rux (jruj2+2)3=2 Ux (jrUj2+2)3=2\r\r\r\r\r2+Ck3exx eyyk2+2C\r\r\r\r\ruy (jruj2+2)3=2 Uy (jrUj2+2)3=2\r\r\r\r\r2+2Ckexyk2; C.B.SCHONLIEBANDA.BERTOZZI447whereweusedauniversalconstantC0fortheuniformbounds.Moreover,foraxedyand0thefunctionx (x2+y2+2)3=2isLipschitzcontinuouswithconstantL(1=),whichisincreasingasdecreases.ByadditionallyapplyingthetriangularinequalityoncemoreweeventuallyhavekH12(ux;uy;uxx;uxy;uyy) H12(Ux;Uy;Uxx;Uxy;Uyy)k2CL(1=)(kexk2+keyk2+krek2)+C(kexxk2+kexyk2+keyyk2);andhencethatH12isLipschitzcontinuous.SimilarlyonecanshowthattheotherH'sareLipschitzcontinuous.LetusfurtherobservethatH1;H4;H5;H6areuniformlyboundedfor0.Moreover,theuniformboundednessofH2andH3forthediscreteintimesolutionUonanitetimeintervalisgivenbythesmoothnessassumptioninTheorem4.1(iii)forU.Then,withtheLipschitzcontinuityandtheuniformboundednessoftheH'sonanitetimeinterval,andtheuniformboundednessonanitetimeintervalofruk,uk,andrukfortheexactsolutionukgiveninTheorem4.1(iii),weeventuallyobtainanestimatefor(4.10):*r0@r0@rUk q jrUkj2+21A r0@ruk q jrukj2+21A1A;rek+1+2C 2krek22+CL3 krek22+C 2krexk22+C 2kreyk22+CL (kexxk22+kexyk22+keyyk22)+C 2krexxk22+C 2krexyk22+C 2kreyyk22+6krek+1k22;(4.11)whereL=L(1=)denotesauniversalLipschitzconstantfortheH'sandCisauniversalconstantfortheinvolveduniformbounds.Further,havingassumedzeroNeumannboundaryconditionsfor(1.5)and(4.4),i.e.,ru~n=rr ru p jruj2+2!~n=0;on@\n;where~nistheoutwardpointingnormalon@\n,thesecondandthirdderivativesin(4.11)canbeboundedbykexxk22+kexyk22+keyyk22+krexxk22+krexyk22+kreyyk22B(kek22+krek22);(4.12)forasuitableconstantB0.BecauseoftheNeumannboundaryconditionswealsogetthatR\ne=0.Hence,wecanapplyPoincare'sinequalitytokek2andobtain,for(4.9),1 2tC2(12)3krek+1k22 C1(11)46krek+1k221 2tC2 220 33CL krekk22C1 1C 2BC5 2L1 krekk221 4kkk2 1;wherewereintroducedtheindexnotationekfore.ThereforebyfollowingthelinesoftheproofofProposition3.4wenallyhave,forktT,krekk22+tM1krekk22T M2eM3T(t)2; 448UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGforsuitablepositiveconstantsM1;M2andM3. Remark4.5.NotethattheLipschitzcontinuityoftheH's{necessaryfortheestimatesintheconvergenceproof{breaksdownif0,whereisthesmoothingparameterinthesquare-rootregularization(4.3)ofthetotalvariation.4.2.Numericalresults.NumericalresultsfortheTV-H 1inpaintingap-proacharepresentedinFigures4.1and4.2.ForacomparisonofthehigherorderTV-H 1inpaintingapproachwithitssecondordercousin,thestandardTV-L2in-paintingmethod,inFigure4.2weconsidertheperformanceofbothalgorithmsinasmallpartoftheimageinFigure4.1.InfacttheresultshowninFigures4.1and4.2stronglyindicatesthecontinuationofthegradientoftheimagefunctionintotheinpaintingdomain.Arigorousproofofthisobservation,astheoneforCahn-Hilliardinpainting(cf.Section3),isamatteroffutureresearch.Inbothexamplesthetotalvariationjrujisapproximatedby jruj2+andthetimestepsizetischosentobeequaltoone.ThecomputationaltimefortheexampleinFigure4.1isoftheorderof100secondsona1.86GHzprocessorwith1GBRAM. Fig.4.1.TV-H 1inpainting:u(1000)with0=103 Fig.4.2.(l.)u(1000)withTV-H 1inpainting,(r.)u(5000)withTV-L2inpainting5.LCISinpaintingOurlastexamplefortheapplicabilityoftheconvexitysplittingmethodtohigher-orderinpaintingapproachesisinpaintingwithLCIS(1.6).Withf2L2(\n)ourin-paintedimageuevolvesintimeasu= (arctan(u))+(f u):Incontrasttotheothertwoinpaintingmethodsthatwediscussed,thisinpainting C.B.SCHONLIEBANDA.BERTOZZI449equationisagradient\rowinL2fortheenergyE(u)=Z\nG(u)dx+1 2Z\n(f u)2;withG0(y)=arctan(y).ThereforeEyre'sresultinTheorem2.3canbeapplieddirectly.ThefunctionalE(u)issplitintoEc EwithEc(u)=Z\nC1 2(u)2dx+1 2Z\nC2 2juj2dx;E(u)=Z\n G(u)+C1 2(u)2dx+1 2Z\n (f u)2+C2 2juj2dx:Theresultingtime-steppingschemeisUk+1 Uk t+C12Uk+1+C2Uk+1= (arctan(Uk))+C12Uk+(f Uk)+C2Uk:(5.1)AgainweimposehomogeneousNeumannboundaryconditions,useDCTtosolve(5.1),andchoosetheconstantsC1andC2suchthatEcandEareallstrictlyconvexandcondition(2.7)issatised.ThefunctionalEcisconvexforallC1;C20.TherstterminEisconvexifC11.Thisfollowsfromitssecondvariation,namelyr2E1(u)(v;w)=d dsZ(C1(u+sw) arctan((u+sw)))vdxs=0=ZC1 1 1+(u)2vwdx:ForE1tobeconvex,r2E1(u)(v;w)mustbe0forallv;w2C1,andthereforeC1 1 1+(u)20:Substitutings=uweobtainC11 1+s28s2R:Thisinequalityisfullledforalls2RifC11.WeobtainthesameconditiononC1forG0(s)=arctan(s ).FortheconvexityofthesecondtermofE,thesecondconstanthastofulllC20;cf.thecomputationforthettingterminSection4.WiththesechoicesofC1andC2alsocondition(2.7)ofTheorem2.3isautomaticallysatised.5.1.Rigorousestimatesforthescheme.Finallywepresentrigorousresultsfor(5.1).IncontrasttotheinpaintingEquations(1.4)and(1.5),inpaintingwithLCISfollowsavariationalprinciple.Hence,bychoosingtheconstantsC1andC2appropriately,i.e.,C11,C20(cf.thecomputationsabove),Theorem2.3ensuresthattheiterativescheme(5.1)isunconditionallygradientstable.Inadditiontothisproperty,wepresentsimilarresultsasbeforeforCahn-HilliardandTV-H 1inpainting.Theorem5.1.Letubetheexactsolutionof(1.6)anduk=u(kt)theexactsolutionattimekt,foratimestept0andk2N.LetUkbethekthiterateof(5.1)withconstantsC11,C20.Thenthefollowingstatementshold: 450UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTING(i)Undertheassumptionthatkuk 1andkruk2arebounded,thenumericalscheme(5.1)isconsistentwiththecontinuousEquation(1.6)andoforderoneintime.(ii)ThesolutionsequenceUkisboundedonanitetimeinterval[0;T]forallt0.Inparticular,forktT,T-139;.942;0xed,wehavekrUkk22+tK1krUkk22eK2T krU0k22+tK1krU0k22+tTC(\n;D;0;f)(5.2)forsuitableconstantsK1,K2,andaconstantCdependingon\n;D;0;fonly.(iii)Letek=uk Uk.Ifkrukk22K;foraconstantK-139;.942;0;andforallktT(5.3)thentheerrorekconvergestozeroast0.Inparticular,forktT,T-139;.942;0xed,wehavekrekk22+tM1krekk22T M2eM3T(t)2;(5.4)forsuitablenonnegativeconstantsM1;M2andM3.Remark5.1.AsinTheorem4.1(cf.alsoRemark4.1)theconvergenceoftheiteratesUktotheexactsolutionisprovenunderanassumptionontheexactsolution,i.e.,assumption(5.3),whosevalidityisunknowningeneral.However,previousresultsin[15]forthedenoisingcase,i.e.,for(x)=0inallof\n,andforsmoothinitialdataandsmoothf,suggesttheassumptionisalsoreasonablefortheinpaintingcase.TheproofofTheorem5.1isorganizedinthefollowingthreePropositions5.2-5.4.SincetheproofofconsistencyfollowsthelinesofProposition3.2andProposition4.2,wejuststatetheresult.Proposition5.2(Consistency(i)).UnderthesameassumptionsasinTheo-rem5.1andinparticularassumingthatkuk 1andkruk2arebounded,wehavekkk 1=O(t)fort0:Nextwewouldliketoshowtheboundednessofasolutionof(5.1)inthefollowingproposition.Proposition5.3.(Unconditionalstability(ii))UnderthesameassumptionsasinTheorem5.1thesolutionsequenceUkfullls(5.2).Thisgivesboundednessofthesolutionsequenceon[0;T].Proof.Ifwemultiply(5.1)with Uk+1andintegrateover\n,weobtain1 tkrUk+1k22 hrUk;rUk+1i2+C2krUk+1k22+C1krUk+1k22=harctan(Uk);Uk+1i2+C1hrUk;rUk+1i2+hr((f Uk));rUk+1i2+C2hrUk;rUk+1i2: C.B.SCHONLIEBANDA.BERTOZZI451UsingthesameargumentsasintheproofsofProposition3.3and4.3weobtain1 2tkrUk+1k22 krUkk22+C2krUk+1k22+C1krUk+1k22harctan(Uk);Uk+1i2+C1 1krUkk22+C11krUk+1k22+20 22krUkk22+2krUk+1k22+C2 3krUkk22+C23krUk+1k22+C(\n;D;0;f):Now,thersttermontherightsideoftheinequalitycanbeestimatedasfollowsharctan(Uk);Uk+1i2= hrarctan(Uk);rUk+1i2= 1 1+(Uk)2rUk;rUk+121 4\r\r\r\r1 1+(Uk)2rUk\r\r\r\r22+4krUk+1k221 4krUkk22+4krUk+1k22:(5.5)Fromthisweget1 2t+C2(1 3) 2krUk+1k22+(C1(1 1) 4)krUk+1k221 2t+20 22+C2 3krUkk22+C1 1+1 4krUkk22+C(\n;D;0;f):AnalogouslytoSection4.1,withCa=1+t(C2 1);C=C1 1;Cc=1+2t(20+2C2);Cd=4(C1+1);weobtainkrUkk22+tCb CakrUkk22eKTkrU0k22+tCb CakrU0k22+tT2 CaC(\n;D;0;f);whichgivesboundednessofthesolutionsequenceon[0;T]foranyT0andanyt0. Theconvergenceofthediscretesolutiontothecontinuousoneast0isveriedinthefollowingproposition.Proposition5.4(Convergence(iii)).UnderthesameassumptionsasinTheorem5.1andinparticularunderassumption(5.3),theerrorekfullls(5.4).Proof.Sinceallthecomputationsintheconvergenceprooffor(5.1)arethesameasinSection4.1for(4.4)exceptoftheestimatefortheregularizer(arctan(u)),weonlygivethedetailsforthelatterandleavetheresttothereader.Thus,fortheinnerproductinvolvingtheregularizerof(5.1)withintheconvergenceproof,weobtainh (arctan(Uk) arctan(uk));ek+1i2=hr(arctan(Uk) arctan(uk));rek+1i2=hw(Uk)rUk w(uk)ruk;rek+1i2= hw(Uk)rek;rek+1i2 h(w(Uk) w(uk))ruk;rek+1i21 2kw(Uk)jrekjk22+1 21k(w(uk) w(Uk))jrukjk22++1 2krek+1k22; 452UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGwherewehaveusedthatr(arctan(u))=1 1+juj2ru=w(u)ru:Usingtheuniformboundednessofw(s)foralls2R,theuniformboundonrukfromAssumption(5.3),andtheLipschitzcontinuityofw,wegeth (arctan(Uk) arctan(uk));ek+1i2C 2krekk22+CL 21kekk22++1 2krek+1k22:Moreover,becauseofthezeroNeumannboundaryconditionsfullledbysolutionsof(1.6)and(5.1),i.e.,ru~n=r(arctan(u))~n=0;on@\n;where~nistheoutwardpointingnormalon@\n,ekhaszeromeanandwecanapplyPoincare'sinequalitytoobtainh (arctan(Uk)arctan(uk));ek+1i2C 2CL 21krekk221 2krek+1k22:FollowingthesamestepsasintheproofofProposition4.4wenallyhave,forktT,krekk22+tM1krekk22T M2eM3T(t)2forsuitablepositiveconstantsM1;M2andM3. 5.2.Numericalresults.ForthecomparisonwithTV-H 1inpaintingweapply(5.1)tothesameimageasinSection4.2.ThisexampleispresentedinFigure5.1.InFigure5.2theLCISinpaintingresultiscomparedwithTV-H 1-andTV-L2inpainting,forasmallpartinthegivenimage.Againtheresultofthiscomparisonindicatesthecontinuationofthegradientoftheimagefunctionintotheinpaintingdomainforthetwohigher-ordermethods.Arigorousproofofthisobservationisamatteroffutureresearch.Forthenumericalcomputationof(5.1)thearctan(s)wasregularizedbyarctan(s=),0andtchosentobeequalto0:01.TheinpaintedimageinFigure5.1hasbeencomputedinabout90secondsona1.86GHzprocessorwith1GBRAM.6.ConclusionInthispaperwepresentseveralhigherorderPDE-basedmethodsforimagein-painting,alongwithunconditionallystabletime-steppingschemesforthesolutionoftheseequations.SpecicexamplesdiscussedincludeCahn-Hilliardinpainting,TV-H 1inpainting,andinpaintingwithLCIS.Theconstructionoftheseschemesisbasedontheideaofconvexitysplitting,alsointroducedinthispaper.Westudythenu-mericalanalysisoftheschemesincludingconsistency,unconditionalstability,andconvergence.Belowweconsidersomeopenproblemsforthisclassofmethods. C.B.SCHONLIEBANDA.BERTOZZI453 Fig.5.1.LCISinpaintingu(500)with=0:1and0=102. Fig.5.2.(l.)u(1000)withLCISinpainting,(m.)u(1000)withTV-H 1inpainting,(r.)u(5000)withTV-L2inpaintingTheadvantageoffourthorderinpaintingmodels,overmodelsofseconddif-ferentialorder,isthesmoothcontinuationofimagecontents,includingdirec-tionofedges,acrossgapsintheimage.FourthorderPDEsrequireanextraboundaryconditioncomparedwithsecondorderequationsandthisisthemo-tivationforadditionalgeometriccontentprovidedbysuchmethods.However,ingeneral,theadditionalboundaryconditioncouldinvolveanyofthehigherderivatives,andforinpaintingisitdesirabletocontinuetherstderivativeaccrosstheinpaintingregion.Themethodsproposedhereareglobalmeth-odsbasedonanL2delitytermassociatedwiththeknowninformation.ForthespecialcaseoftheCahn-Hilliardequation[14],inthelimitas0!1astationarysolutionisprovedtosatisfypreciselythedesiredtwoboundaryconditions|matchingofgreyvalueandmatchingdirectionofedges.Weconjecturethatanalogousresultsaretruefortheothermethodspresentedherealthougharigorousproofisbeyondthescopeofthismanuscript.Fortheproofsofconvergenceofthediscretesolutiontotheexactsolution,i.e.,fortheproofsofTheorem4.4andTheorem5.4,wehadtoassumethattheexactsolutionisboundedonanitetimeintervalinacertainSobolevnorm.Aswealreadyarguedintheremarksafterthestatementofthetheorems,theseassumptionsseemtobeheuristicallyreasonableconsideringearlierre-sultsin[15,16].Neverthelessarigorousderivationofsuchboundsisstillmissing.Besidesthefactthatrigorousresultsforfourth-orderpartialdierentialequa-tionsarerareingeneral,anasymptoticanalysisofourthreeinpaintingmodelswouldbeofhigh(evenpractical)interest.MorepreciselytheconvergenceofasolutionoftheevolutionEquations(1.4),(1.5),and(1.6),toastation-arystateisstillopen.Sincetheinpaintedimageisthestationarysolution 454UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGofthoseevolutionequations,theasymptoticbehaviorisofcourseanissue.Also,inpracticethenumericalschemesaresolvedtosteadystate(uptoanapproximationalerror).Notethatinadditiontothefourthdierentialorder,adicultyintheconvergenceanalysisof(1.4)and(1.5)isthattheequationsdonotfollowavariationalprinciple.Thediscreteschemesproposedinthispaperareunconditionallystableandtheirnumericalperformanceisamatterof10to100secondsforsmalltomedium-sizedimages,i.e.,128128to256256pixels,andgapsthatcon-stituteaboutonetotenpercentoftheimagedomain.Fastnumericalsolversforhigherorderinpaintingmodelsisstillamostlyopeneldofresearch.AmongsuchfastsolverswefoundtherecentcontributionofBrito-LoezaandChen[18]veryinterestingandforward-looking,whouseamultigridmethodtosolveinpaintingwithCDD(CurvatureDrivenDiusion).Anotherap-proachistheSplitBregmanmethodofGoldsteinandOsher[41,42],whichsuggestsasplittingofahigher-ordervariationalproblemintwoconsecutivelyminimizedrst-orderproblems.Althoughnotdirectlyapplicabletothenon-variationalinpaintingtechniques(1.4)and(1.5),theirmethodpromisesanecientsolutionof,e.g.,(1.6)LCISinpainting.Acknowledgments.C.-B.Schonliebacknowledgesthenancialsupportpro-videdbytheDFGGraduiertenkolleg1023IdenticationinMathematicalModels:SynergyofStochasticandNumericalMethods,supportbytheprojectWWTFFivesenses-Call2006,MathematicalMethodsforImageAnalysisandProcessingintheVi-sualArtsandbytheFFGprojectErarbeitungneuerAlgorithmenzumImageInpaint-ing,projectnumber813610.Further,thispublicationisbasedonworksupportedbyAwardNo.KUK-I1-007-43,madebyKingAbdullahUniversityofScienceandTech-nology(KAUST).Inaddition,C.-B.SchonliebthanksIPAM(InstituteforPureandAppliedMathematics),UCLA,forthehospitalityandthenancialsupportduringthepreparationofthiswork.BothauthorsacknowledgesupportfromtheNSFgrantBCS-0527388,ONRgrantN000140810363,andtheDepartmentofDefense.TheauthorsthanktherefereesandWenuahGao(UCLA)forusefulcommentsonthemanuscript.REFERENCES[1]J.-F.AujolandA.Chambolle,Dualnormsandimagedecompositionmodels,InternationalJournalofComputerVision,63(1),85{104,June,2005.[2]J.-F.AujolandG.Gilboa,ConstrainedandSNR-basedsolutionsforTV-Hilbertspaceimagedenoising,J.Math.ImagingandVision,26(1-2),217{237,November2006.[3]W.Baatz,M.Fornasier,P.MarkowichandC.B.Schonlieb,InpaintingofancientAustrianfrescoes,ConferenceProceedingsofBridges2008,Leeuwarden2008,150{156,2008.[4]J.W.BarrettandJ.F.Blowey,Finiteelementapproximationofamodelforphaseseparationofamulti-componentalloywithnon-smoothfreeenergy,Numer.Math.,77(1),1{34,1997.[5]J.W.BarrettandJ.F.Blowey,Finiteelementapproximationofamodelforphaseseparationofamulti-componentalloywithaconcentration-dependentmobilitymatrix,IMAJ.Numer.Anal.,18(2),287{328,1998.[6]J.W.BarrettandJ.F.Blowey,Finiteelementapproximationofamodelforphaseseparationofamulti-componentalloywithnonsmoothfreeenergyandaconcentrationdependentmobilitymatrix,Math.ModelsMethodsAppl.Sci.,9(5),627{663,1999.[7]J.W.BarrettandJ.F.Blowey,FiniteelementapproximationoftheCahn-Hilliardequationwithconcentrationdependentmobility,Math.Comput.,68(226),487{517,1999. C.B.SCHONLIEBANDA.BERTOZZI455[8]J.Becker,G.Grun,M.LenzandM.Rumpf,Numericalmethodsforfourthordernonlineardiusionproblems,ApplicationsofMathematics,47,517{543,2002.[9]J.Bect,L.Blanc-Feraud,G.AubertandA.Chambolle,A`1-uniedVariationalFrameworkforImageRestoration,T.PajdlaandJ.Matas(eds.),Proceedingsofthe8thEuropeanConferenceonComputerVisionIV,Prague,CzechRepublic,SpingerVerlag,2004.[10]M.Bertalmio,G.Sapiro,V.CasellesandC.Ballester,Imageinpainting,Siggraph2000,Com-puterGraphicsProceedings,417{424,2000.[11]M.Bertalmio,A.L.BertozziandG.Sapiro,Navier-Stokes,\ruiddynamics,andimageandvideoinpainting,Proceedingsofthe2001IEEEComputerSocietyConferenceonComputerVisionandPatternRecognition,1,355{362,2001.[12]M.Bertalmio,L.Vese,G.SapiroandS.Osher,Simultaneousstructureandtextureimageinpainting,IEEETrans.ImageProc.,12(8),882{889,2003.[13]A.Bertozzi,S.EsedogluandA.Gillette,InpaintingofbinaryimagesusingtheCahn-Hilliardequation,IEEETrans.ImageProc.,16(1),285{291,2007.[14]A.Bertozzi,S.EsedogluandA.Gillette,Analysisofatwo-scaleCahn-Hilliardmodelforimageinpainting,MultiscaleModelingandSimulation,6(3),913{936,2007.[15]A.BertozziandJ.B.Greer,Low-curvatureimagesimpliers:globalregularityofsmoothsolu-tionsandLaplacianlimitingschemes,Commun.PureAppl.Math.,LVII,764{790,2004.[16]A.Bertozzi,J.Greer,S.OsherandK.Vixie,NonlinearregularizationsofTVbasedPDEsforimageprocessing,AMSSeriesofContemporaryMathematics,G.-Q.Chen,G.Gasper,andJ.Jerome(eds.),371,29{40,2005.[17]A.Bertozzi,N.JuandH.W.Lu,Abiharmonic-modiedforwardtimesteppingmethodforfourthordernonlineardiusionequations,DiscreteandContinuousDynamicalSystems2009,submitted.UCLACAMreports08-18,2008.[18]C.Brito-LoezaandK.Chen,Multigridmethodforamodiedcurvaturedrivendiusionmodelforimageinpainting,J.Comput.Math.,26(6),856{875,2008.[19]M.Burger,L.HeandC.Schonlieb,Cahn-Hilliardinpaintingandageneralizationforgrayvalueimages,SIAMJ.ImagingSci.,2(4),1129{1167,2009.[20]F.Cao,Y.Gousseau,S.MasnouandP.Perez,Geometricallyguidedexemplar-basedinpainting,INRIApreprint,hal-00380394,2009..[21]V.Caselles,J.M.Morel,C.SbertandA.Gillette,Anaxiomaticapproachtoimageinterpola-tion,IEEETrans.onImageProc.,7(3),376{386,1998.[22]A.Chambolle,Analgorithmfortotalvariationminimizationandapplications,J.Math.Imag-ingVis.,20(1{2),89{97,2004.[23]T.F.ChanandJ.Shen,Mathematicalmodelsforlocalnon-textureinpaintings,SIAMJ.Appl.Math.,62(3),1019{1043,2001.[24]T.F.ChanandJ.Shen,Non-textureinpaintingbycurvaturedrivendiusions(CDD),J.VisualCommun.ImageRep.,12(4),436{449,2001.[25]T.F.ChanandJ.Shen,Variationalrestorationofnon-\ratimagefeatures:modelsandalgo-rithms,SIAMJ.Appl.Math.,61(4),1338{1361,2001.[26]T.F.ChanandJ.Shen,Variationalimageinpainting,Commun.PureAppliedMath.,58,579{619,2005.[27]T.F.Chan,S.H.KangandJ.Shen,Euler'selasticaandcurvature-basedinpainting,SIAMJ.Appl.Math.,63(2),564{592,2002.[28]T.F.Chan,J.H.ShenandH.M.Zhou,Totalvariationwaveletinpainting,J.Math.ImagingVis.,25(1),107{125,2006.[29]A.Criminisi,P.PerezandK.Toyama,Objectremovalbyexemplar-basedinpainting,IEEEInt.Conf.Comp.VisionandPatternRecog.2,721{728,2003.[30]J.A.DobrosotskayaandA.L.Bertozzi,AWavelet-laplacevariationaltechniqueforimagede-convolutionandinpainting,IEEETrans.Imag.Proc.,17(5),657{663,2008.[31]J.Jr.DouglasandT.Dupont,Alternating-directionGalerkinmethodsonrectanglesnumericalsolutionofpartialdierentialequations,II(SYNSPADE1970)(Proc.Symp.,UniversityofMaryland,CollegePark,Md.1970)(NewYork:Academic),133{214,1971.[32]A.A.EfrosandT.K.Leung,TextureSynthesisbyNon-parametricSampling,IEEEInterna-tionalConferenceonComputerVision,Corfu,Greece,September,1999.[33]C.M.ElliottandS.A.Smitheman,AnalysisoftheTVregularizationandH 1delitymodelfordecomposinganimageintocartoonplustexture,Commun.PureAppl.Anal.,6(4),917{936,2007.[34]C.M.ElliottandS.A.Smitheman,NumericalanalysisoftheTVregularizationandH 1delitymodelfordecomposinganimageintocartoonplustexture,IMAJournalofNumericalAnalysis,1{39,2008.[35]S.EsedogluandJ.H.Shen,DigitalinpaintingbasedontheMumford-Shah-Eulerimagemodel, 456UNCONDITIONALLYSTABLESCHEMESFORHIGHERORDERINPAINTINGEur.J.Appl.Math.,13(4),353{370,2002.[36]D.Eyre,Anunconditionallystableone-stepschemeforgradientsystems,Jun.,1998,unpub-lished.[37]X.FengandA.Prohl,ErroranalysisofamixedniteelementmethodfortheCahn-Hilliardequation,Numer.Math.,99(1),47{84,2004.[38]X.FengandA.Prohl,NumericalanalysisoftheCahn-HilliardequationandapproximationfortheHele-Shawproblem,InterfacesFreeBound.,7(1),1{28,2005.[39]M.FornasierandC.B.Schonlieb,Subspacecorrectionmethodsfortotalvariationandl1-min-imization,SIAMJ.Numer.Anal.,47(5),3397{3428,2009.[40]K.Glasner,AdiuseinterfaceapproachtoHele-Shaw\row,Nonlinearity,16,1{18,2003.[41]T.Goldstein,X.BressonandS.Osher,GeometricapplicationsofthesplitBregmanmethod:Segmentationandsurfacereconstruction,J.Sci.Comp.,45(1-3),pp.272{293,2010.[42]T.GoldsteinandS.Osher,ThesplitBregmanmethodforL1regularizedproblems,SIAMJournalonImagingSciences,2(2),323{343,2009.[43]J.B.GreerandA.Bertozzi,H1solutionsofaclassoffourthordernonlinearequationsforimageprocessing,DiscreteContin.Dyn.Syst.,10(1-2),349{366,2004.[44]J.B.GreerandA.Bertozzi,TravelingwavesolutionsoffourthorderPDEsforimageprocessing,SIAMJ.Math.Anal.,36(1),38{68,2004.[45]H.GrossauerandO.Scherzer,UsingthecomplexGinzburg-Landauequationfordigitalin-paintingin2Dand3D,ScaleSpaceMethodsinComputerVision,2695,225{236,2003.[46]G.GrunandM.Rumpf,Nonnegativitypreservingconvergentschemesforthethinlmequa-tion,Num.Math.,87,113{152,2000.[47]J.W.Gu,L.Zhang,G.Q.Yu,Y.X.XingandZ.Q.Chen,X-rayCTmetalartifactsreductionthroughcurvaturebasedsinograminpainting,JournalofX-RayScienceandTechnology,14(2),73{82,2006.[48]M.HirschandS.Smale,DierentialEquations,DynamicalSystemsandLinearAlgebra,Aca-demicPress,London,1974.[49]L.LieuandL.Vese,ImagerestorationanddecompostionviaboundedtotalvariationandnegativeHilbert-Sobolevspaces,AppliedMathematics&Optimization,58,167{193,2008.[50]M.Lysaker,A.LundervoldandX.C.Tai,Noiseremovalusingfourth-orderpartialdierentialequationswithapplicationstomedicalmagneticresonanceimagesinspaceandtime,IEEETrans.onImageProc.,12(12),1579{1590,2003.[51]O.M.LysakerandX.-C.Tai,Iterativeimagerestorationcombiningtotalvariationminimiza-tionandasecond-orderfunctional,Inter.J.Comput.Vis.,66(1),5{18,2006.[52]S.MasnouandJ.Morel,Levellinesbaseddisocclusion,5thIEEEInt'lConf.onImagePro-cessing,Chicago,IL,Oct.,4-7,259{263,1998.[53]L.ModicaandS.Mortola,Illimitenella -convergenzadiunafamigliadifunzionaliellittici,Boll.UnioneMat.Ital.,V.Ser.,A14,526{529,1977.[54]L.ModicaandS.Mortola,Unesempiodi -convergenza,Boll.UnioneMat.Ital.,V.Ser.,B14,285{299,1977.[55]M.J.NarashimaandA.M.Peterson,Onthecomputationofthediscretecosinetransform,IEEETrans.Commun.COM,26,934-936,June,1978.[56]L.K.Nielsen,X.C.Tai,S.I.AanonsenandM.Espedal,NoiseRemovalofSeismicDataUsingaFourth-OrderParabolicPDE,InternationalJ.Tomography&Statistics(IJTS),4(6),63,2006.[57]M.Nitzberg,D.MumfordandT.Shiota,Filtering,Segmentation,andDepth,Springer-Verlag,LectureNotesinComputerScience,662,1993.[58]S.Osher,A.SoleandL.Vese,ImagedecompositionandrestorationusingtotalvariationminimizationandtheH 1norm,MultiscaleModelingandSimulation:ASIAMInterdis-ciplinaryJournal,1(3),349{370,2003.[59]P.PeronaandJ.Malik,Scale-spaceandEdgeDetectionUsingAnisotropicDiusion,IEEETransactionsonPatternAnalysisandMachineIntelligence,12(7),629{639,July,1990.[60]L.I.Rudin,S.OsherandE.Fatemi,Nonlineartotalvariationbasednoiseremovalalgorithms,PhysicaD,60,259{268,1992.[61]L.RudinandS.Osher,Totalvariationbasedimagerestorationwithfreelocalconstraints,Proc.1stIEEEICIP,1,31{35,1994.[62]C.B.Schonlieb,TotalvariationminimizationwithanH 1constraint,CRMSeries9,Singu-laritiesinNonlinearEvolutionPhenomenaandApplicationsProceedings,ScuolaNormaleSuperiorePisa,2009,201{232.[63]C.B.Schonlieb,ModernPDETechniquesforImageInpainting,PhDThesis,UniversityofCambridge,265,June2009.[64]C.B.Schonlieb,A.Bertozzi,M.BurgerandL.He,ImageInpaintingUsingaFourth-Order C.B.SCHONLIEBANDA.BERTOZZI457TotalVariationFlow,Proc.Int.Conf.SampTA09,Marseilles,2009.[65]P.Smereka,Semi-implicitlevelsetmethodsforcurvatureandsurfacediusionmotion,SpecialissueinhonorofthesixtiethbirthdayofStanleyOsher,J.Sci.Comput.,19(1-3),439{456,2003.[66]A.M.StuartandA.R.Humphries,Modelproblemsinnumericalstabilitytheoryforinitialvalueproblems,SIAMRev.,36(2),226{257,1994.[67]R.Temam,Innite-DimensionalDynamicalSystemsinMechanicsandPhysics,Appl.Math.Sci.,Springer,NewYork,68,1988.[68]A.Tsai,J.A.YezziandA.S.Willsky,CurveevolutionimplementationoftheMumford-Shahfunctionalforimagesegmentation,denoising,interpolationandmagnication,IEEETrans.ImageProcess,10(8),1169{1186,2001.[69]J.TumblinandG.Turk,LCIS:ABoundaryhierarchyfordetail-preservingcontrastreduction,Siggraph,ComputerGraphicsProceedings,83{90,1999.[70]L.VeseandS.Osher,Modelingtextureswithtotalvariationminimizationandoscillatingpatternsinimageprocessing,J.Sci.Comput.,19(1-3),553{572,2003.[71]B.P.Vollmayr-LeeandA.D.Rutenberg,Fastandaccuratecoarseningsimulationwithanun-conditionallystabletimestep,Phys.Rev.E,68(0066703),1{13,2003.[72]L.Y.WeiandM.Levoy,Fasttexturesynthesisusingtree-structuredvectorquantization,InProc.SIGGRAPH'00,NewOrleans,USA,479{488,2000.[73]A.L.YuilleandA.Rangarajan,TheConcave-Convexprocedure(CCCP),NeuralComputation,15(4),915{936,April,2003.