數(shù)字視頻特效處理lec04-tonemapping色調(diào)映射

TonemappingDigitalVisualEffectsYung-YuChuangwithslidesbyFredoDurand,Lin-YuTseng,andAlexeiEfrosTonemappingHowshouldwemapsceneluminances(upto1:100,000)todisplayluminances(onlyaround1:100)toproduceasatisfactoryimage?Linearscaling?,thresholding?10-610610-6106RealworldradianceDisplayintensitydynamicrangePixelvalue0to255CRThas300:1dynamicrangeTheultimategoalisavisualmatchWedonotneedtoreproducethetrueradianceaslongasitgivesusavisualmatch.visualadaptionEyeisnotaphotometer!Dynamicrangealongthevisualpathwayisonlyaround32:1.ThekeyisadaptationEyeisnotaphotometer!Aretheheadlightsdifferentintwoimages?Physically,theyarethesame,butperceptuallydifferent.WearemoresensitivetocontrastWeber’slawbackgroundintensityJust-noticeableDifference(JND)flashHowhumansdealwithdynamicrangeWe'remoresensitivetocontrast(multiplicative)Aratioof1:2isperceivedasthesamecontrastasaratioof100to200MakessensebecauseilluminationhasamultiplicativeeffectUsethelogdomainasmuchaspossibleDynamicadaptation(verylocalinretina)Pupil(notsoimportant)NeuralChemicalDifferentsensitivitytospatialfrequenciesPreliminariesForcolorimagesLogdomainisusuallypreferred.HDRDisplayOncewehaveHDRimages(eithercapturedorsynthesized),howcanwedisplaythemonnormaldisplays?HDRdisplaysystem,SunnybrookTechnology,SIGGRAPH2004DLP800:1LCD300:1diffuserTheoretically,240,000:1.Duetoimperfectopticaldepth,54,000:1measuredSunnybrookHDRdisplaySlidefromthe2005SiggraphcourseonHDRHowitworksSlidefromthe2005SiggraphcourseonHDRBrightsideHDRdisplay37”200000:1AcquiredbyDolbyTonemappingoperatorsSpatial(global/local)FrequencydomainGradientdomain3papersfromSIGGRAPH2002PhotographicToneReproductionforDigitalImagesFastBilateralFilteringfortheDisplayofHigh-Dynamic-RangeImagesGradientDomainHighDynamicRangeCompressionPhotographicToneReproductionforDigitalImagesErikReinhardMikeStarkPeterShirleyJimFerwerdaSIGGRAPH2002Globalv.s.localPhotographictonereproductionProposedbyReinhardet.al.inSIGGRAPH2002Motivatedbytraditionalpractice,zonesystembyAnselAdamsanddodgingandburningItcontainsbothglobalandlocaloperatorsZonesystemTheZonesystemFormalismtotalkaboutexposure,densityZone=intensityrange,inpowersoftwoInthescene,onthenegative,ontheprintSource:AnselAdamsTheZonesTheZonesystemYoudecidetoputpartofthesysteminagivenzoneDecision:exposure,development,printDodgingandburningDuringtheprintHidepartoftheprintduringexposureMakesitbrighterFromTheMasterPrintingCourse,RudmanDodgingandburningFromPhotographybyLondonetal.dodgingburningDodgingandburningMustbedoneforeverysingleprint!StraightprintAfterdodgingandburningGlobaloperatorApproximationofscene’skey(howlightordarkitis).Mapto18%ofdisplayrangeforaverage-keysceneUser-specified;highkeyorlowkeytransferfunctiontocompresshighluminancesGlobaloperatorItseldomreaches1sincetheinputimagedoesnothaveinfinitelylargeluminancevalues.Lwhite

isthesmallestluminancetobemappedto1lowkey(0.18)highkey(0.5)Dodgingandburning(localoperators)AreareceivingadifferentexposureisoftenboundedbysharpcontrastFindlargestsurroundingareawithoutanysharpcontrastDodgingandburning(localoperators)Adarkerpixel(smallerthantheblurredaverageofitssurroundingarea)isdividedbyalargernumberandbecomedarker(dodging)Abrighterpixel(largerthantheblurredaverageofitssurroundingarea)isdividedbyasmallernumberandbecomebrighter(burning)BothincreasethecontrastDodgingandburningFrequencydomainFirstproposedbyOppenheimin1968!Undersimplifiedassumptions,image=illuminance*reflectancelow-frequencyattenuatemorehigh-frequencyattenuatelessOppenheimTakingthelogarithmtoformdensityimagePerformFFTonthedensityimageApplyfrequency-dependentattenuationfilterPerforminverseFFTTakeexponentialtoformthefinalimageFastBilateralFilteringfortheDisplayofHigh-Dynamic-RangeImagesFrédoDurand&JulieDorseySIGGRAPH2002AtypicalphotoSunisoverexposedForegroundisunderexposedGammacompressionX->XgColorsarewashed-outInputGammaGammacompressiononintensityColorsareOK,butdetails(intensityhigh-frequency)areblurredGammaonintensityIntensityColorChiuetal.1993Reducecontrastoflow-frequenciesKeephighfrequenciesReducelowfrequencyLow-freq.High-freq.ColorThehalonightmareForstrongedgesBecausetheycontainhighfrequencyReducelowfrequencyLow-freq.High-freq.ColorDurandandDorseyDonotbluracrossedgesNon-linearfilteringOutputLarge-scaleDetailColorEdge-preservingfilteringBlur,butnotacrossedgesAnisotropicdiffusion[Perona&Malik90]BlurringasheatflowLCIS[Tumblin&Turk]Bilateralfiltering[Tomasi&Manduci,98]Edge-preservingGaussianblurInputStartwithGaussianfilteringHere,inputisastepfunction+noiseoutputinputStartwithGaussianfilteringSpatialGaussianfoutputinputStartwithGaussianfilteringOutputisblurredoutputinputGaussianfilterasweightedaverageWeightofxdependsondistancetoxoutputinputTheproblemofedgesHere,“pollutes”ourestimateJ(x)ItistoodifferentoutputinputPrincipleofBilateralfiltering[TomasiandManduchi1998]PenaltygontheintensitydifferenceoutputinputBilateralfiltering[TomasiandManduchi1998]SpatialGaussianfoutputinputBilateralfiltering[TomasiandManduchi1998]SpatialGaussianfGaussiangontheintensitydifferenceoutputinputoutputinputNormalizationfactor[TomasiandManduchi1998]k(x)=outputinputBilateralfilteringisnon-linear[TomasiandManduchi1998]TheweightsaredifferentforeachoutputpixelContrastreductionInputHDRimageContrasttoohigh!ContrastreductionColorInputHDRimageIntensityContrastreductionColorIntensityLargescaleFast

Bilateral

FilterInputHDRimageContrast

reductionDetailColorIntensityLargescaleFast

Bilateral

FilterInputHDRimageContrastreductionDetailColorIntensityLargescaleFast

Bilateral

FilterReduce

contrastLargescaleInputHDRimageScaleinlogdomainContrastreductionDetailColorIntensityLargescaleFast

Bilateral

FilterReduce

contrastDetailLarge

scalePreserve!InputHDRimageContrastreductionDetailColorIntensityLargescaleFast

Bilateral

FilterReduce

contrastDetailLarge

scaleColorPreserve!InputHDRimageOutputBilateralfilterisslow!ComparedtoGaussianfiltering,itismuchslowerbecausethekernelisnotfixed.DurandandDorseyproposedanapproximateapproachtospeedupParisandDurandproposedaneven-fasterapproachinECCV2006.Wewillcoverthisonewhentalkingaboutcomputationalphotogrphy.OppenheimbilateralGradientDomainHighDynamicRangeCompressionRaananFattalDaniLischinskiMichaelWermanSIGGRAPH2002LogdomainLogorithmisacrudeapproximationtotheperceivedbrightnessGradientsinlogdomaincorrespondtoratios(localcontrast)intheluminancedomainThemethodin1DlogderivativeattenuateintegrateexpThemethodin2DGiven:alog-luminanceimageH(x,y)Computeanattenuationmap

ComputeanattenuatedgradientfieldG:Problem:G

maynotbeintegrable!SolutionLookforimageIwithgradientclosesttoGintheleastsquaressense.Iminimizestheintegral:PoissonequationSolve..1…1-41…1..ISolvingPoissonequationNoanalyticalsolutionMultigridmethodConjugategradientmethodAttenuationAnydramaticchangeinluminanceresultsinlargeluminancegradientatsomescaleEdgesexistinmultiplescales.Thus,wehavetodetectandattenuatethematmultiplescalesConstructaGaussianpyramidHiAttenuationgradientmagnitudelog(Luminance)attenuationmapMultiscalegradientattenuationinterpolateinterpolateX=X=FinalgradientattenuationmapPerformanceMeasuredon1.8GHzPentium4:

512x384:1.1sec1024x768:4.5secCanbeacceleratedusingprocessor-optimizedlibraries.Bilateral

[Durandetal.]Photographic

[Reinhardetal.]Gradientdomain

[Fattaletal.]InformalcomparisonInformalcomparisonBilateral

[Durandetal.]Photographic

[Reinhardetal.]Gradientdomain

[Fattaletal.]Bilateral

[Durandetal.]Photographic

[Reinhardetal.]Gradientdomain

[Fattaletal.]InformalcomparisonEvaluationofToneMappingOperatorsusingaHighDynamicRangeDisplayPatrickLeddaAlanChalmersTomTroscinkoHelgeSeetzenSIGGRAPH2005SixoperatorsH:histogramadjustmentB:bilateralfilterP:photographicreproductionI:iCAML:logarithmmappingA:localeyeadaption23scenesExperimentsettingHDRdisplaytonemappingresulttonemappingresultPreferencematrixRankingiseasierthanrating.15pairsforeachpersontocompare.Atotalof345pairspersubject.preferencematrix(tmo2->tmo4,tom2isbetterthantmo4)StatisticalmeasurementsStatisticalmeasurementsareusedtoevaluate:Agreement:whethermostagreeontherankingbetweentwotonemappingoperators.Consistency:nocycleinranking.Ifallareconfusedin