何恺明在CVPR会议上演讲时的ppt——基于暗通道先验的单幅图像去雾技术

SingleImageHazeRemovalUsingDarkChannelPriorKaimingHeJianSunXiaoouTangTheChineseUniversityofHongKongMicrosoftResearchAsiaTheChineseUniversityofHongKongHazyImages•Lowvisibility•FaintcolorsGoalsofHazeRemoval•Scenerestoration•DepthestimationdepthHazeImagingModel)1(ttAJITransmissionSceneradianceHazyimageAtmosphericlightTransmissionHazeImagingModel)1(ttAJIDepthtdlnAmbiguityinHazeRemovalinputsceneradiancedepth….PreviousWorks•Usingadditionalinformation–Polarizationfilter[Shwartzetal.,CVPR’06]–Multipleimages[Narasimhan&Nayar,CVPR’00]–Known3Dmodel[Kopfetal.,SiggraphAsia’08]–User-assistance[Narasimhan&Nayar,CPMCV’03]PreviousWorks•Singleimage–Maximizelocalcontrast[Tan,CVPR08]PreviousWorks•Singleimage–Maximizelocalcontrast[Tan,CVPR08]PreviousWorks•Singleimage–Maximizelocalcontrast[Tan,CVPR08]–IndependentComponentAnalysis[Fattal,Siggraph08]PreviousWorks•Singleimage–Maximizelocalcontrast[Tan,CVPR08]–IndependentComponentAnalysis[Fattal,Siggraph08]PriorsinComputerVision•Smoothnessprior•Sparsenessprior•Exemplar-basedpriorIll-posedproblemwell-posedproblempriorDarkChannelPriorDarkChannel•min(rgb,localpatch)DarkChannel•min(rgb,localpatch)–min(r,g,b)min(r,g,b)DarkChannel•min(rgb,localpatch)–min(r,g,b)–min(localpatch)=minfilter15x15darkestdarkchannelDarkChannel•min(rgb,localpatch)–min(r,g,b)–min(localpatch)=minfilterdarkchannel))(Jmin(min)(Jc}bg,r,{c)(yyxxdark–Jc:colorchannelofJ–Jdark:darkchannelofJDarkChannel•min(rgb,localpatch)–min(r,g,b)–min(localpatch)=minfilterdarkchannel)Jmin(minJccdark–Jc:colorchannelofJ–Jdark:darkchannelofJASurprisingObservationHaze-freeASurprisingObservationHaze-freeASurprisingObservationHaze-freeASurprisingObservationHaze-freeASurprisingObservationHaze-freeASurprisingObservationHaze-freeASurprisingObservation00.20.40.60.81064128192256Prob.Pixelintensityofdarkchannels86%pixelsin[0,16]5,000haze-freeimagesDarkChannelPrior•Foroutdoorhaze-freeimages0)Jmin(minccWhatmakesitdark?•Blackobject•Colorfulobject•ShadowDarkChannelofHazyImage•Thedarkchannelisnolongerdark.hazyimagedarkchannelTransmissionEstimation)1(ttAJIHazeimagingmodeltt1AJAIccccNormalizett1)AJmin(min)AImin(minccccccComputedarkchanneltt1)AJmin(min)AImin(minccccccComputedarkchannelTransmissionEstimation0)Jmin(minccDarkChannelPrior0Computedarkchanneltt1)AJmin(min)AImin(minccccccTransmissionEstimationEstimatetransmission)AImin(min1ccctTransmissionEstimationinputItestimated)AImin(min1ccctEstimatetransmissionTransmissionOptimization)1(ttAJIHazeimagingmodel)1(BFIMattingmodel+Refinedtransmission+tri-mapTransmissionOptimization•L-mattingLaplacian[Levinetal.,CVPR‘06]•Constraint-soft,dense(matting-hard,sparse)LtttttT2~)(DatatermSmoothnesstermTransmissionOptimizationbeforeoptimizationTransmissionOptimizationafteroptimizationhazyimagedarkchannelbrightestpixelsAtmosphericLightEstimationbrightestpixelA:mosthazySceneRadianceRestoration)1(ttAJISceneradianceTransmissionHazyimageAtmosphericlightResultsinputResultsrecoveredimageResultsdepthResultsinputResultsrecoveredimageResultsdepthResultsinputResultsrecoveredimageResultsdepthComparisonsinput[FattalSiggraph08]ComparisonsinputourresultComparisonsinput[Tan,CVPR08]Comparisonsinputourresultinputourresult[Kopfetal,SiggraphAsia08]ComparisonsResults:De-focusrecoveredsceneradianceinputdepthinputdepthResults:De-focusde-focusResults:VideooutputinputResults:Videooutputinputinputourresulttransmission•InherentlywhiteorgrayishobjectsLimitations•Hazeimagingmodelisinvalid–e.g.non-constantAinputourresultLimitationsSummary•Darkchannelprior–Anaturalphenomenon–Verysimplebuteffective–PutabadimagetogooduseThankyou