Multiresolution sampling procedure for analysis and synthesis of texture images

JeremyS.DeBonet

TheProblem:

Wearedevelopingnewtechniquesfortreatinginputtextureimagesasprobabilitydensityestimatorsfromwhichwecansamplenewtextures,withsimilarappearanceandstructuralproperties.

Motivation:

Thegoaloftexturesynthesisistogenerateanewimagefromanexampletexture,suchthatthenewimageissufficientlydifferentfromtheoriginalyetstillappearsasthoughitwasgeneratedbythesameunderlyingstochasticprocess.Ifsuccessful,thenewimagewilldifferfromtheoriginal,yethaveperceptuallyidenticaltexturecharacteristics.

PreviousWork:

Recentlyseveralattemptsatdevelopingsuchtechniqueshavebeensuccessfulinlimiteddomains.MostnotablyHeegerandBergen[1],whoiterativelyresamplerandomnoisetocoerceitintohavingparticularmultiresolutionorientedenergyhistograms.

Approach:

Theapproachpresentedhereusestheresultspsychophysicalmodelstoprovideconstraintsonastatisticalsamplingprocedure.Inatwo-phaseprocess,theinputtextureisfirstanalyzedbymeasuringthejointoccurrence,acrossmultipleresolutions,ofseveralofthefeaturesusedinpsychophysicalmodels.Inthesecondphase,anewtextureissynthesizedbysamplingsuccessivespatialfrequencybandsfromtheinputtexture,conditionedonthesimilarjointoccurrenceoffeaturesatalllowerspatialfrequencies.Byrearrangingtexturalcomponentsatlocationsandresolutionswherethediscriminabilityisbelowthreshold,newtexturesamplesaregeneratedwhichhavesimilarvisualcharacteristics.SeveralresultsofthismethodareshowninFigure1.

Difficulty:

Thegoaloftexturesynthesisistogenerateatexturewhichisbothrandomandperceptuallyidenticaltotheoriginaltexture.Todothisthesynthesizedtexturemustappeartohavebeengeneratedbythesameunderlyingphysicalprocessastheoriginal.However,thisprocessisunknownandthesuccessofsynthesisdependsuponitsabilitytodiscoverit.

Impact:

Fromthisresearchnewmodelsofprobabilistictexturegenerationhavedeveloped.Thesesuggestunderlyingimagerepresentationswhichcanbeusedincomputervisionandimagedatabasetechnology.

Futurework:

Newtexturediscriminationtechniques,whichisalargecomponentofgeneralimagerecognition,canbedevelopedasaresultofthisresearch.

ResearchSupport:

ResearchsupportedinpartbyDARPAcontractNo.N00014-94-1-0128.References:

[1]Heeger,D.J.andBergen,J.R.1995.Pyramidbasedtextureanalysis/synthesis.InComputerGraphics,

pages229–238.ACMSIGGRAPH.[2]DeBonet,J.S.1997.Multiresolutionsamplingprocedureforanalysisandsynthesisoftextureimages.

InComputerGraphics.ACMSIGGRAPH.

Figure1:Texturesynthesisresults.Thesmallerpatchesaretheinputtextures,andtotheirrightaresynthesizedimageswhichare4or9timeslarger.