SSIM课件

PresentedBy:BishwaShethStructuralSimilarityIndexGuidedBy:Dr.K.R.RaoTopicstobeCoveredWhyImagequalitymeasureWhatisImagequalitymeasureTypesofqualityassessmentMSE–MeansquareerrorSSIM-StructuralsimilarityindexmethodVIF–VirtualinformationfidelitySimulationresultsConclusionReferences2WhyImagequality?Digitalimagesaresubjecttowidevarietyofdistortionsduringtransmission,acquisition,processing,compression,storageandreproductionanyofwhichmayresultindegradationofvisualqualityofanimage.E.g.lossycompressiontechnique–usedtoreducebandwidth,itmaydegragethequalityduringquantizationprocess.Sotheultimateaimofdatacompressionistoremovetheredundancyfromthesourcesignal.Thereforeitsreducesthenoofbinarybitsrequiredtorepresenttheinformationcontainedwithinthesource.3WhatisImageQualityAssessment?ImagequalityisacharacteristicofanimagethatmeasurestheperceivedimagedegradationItplaysanimportantroleinvariousimageprocessingapplication.Goalofimagequalityassessmentistosupplyqualitymetricsthatcanpredictperceivedimagequalityautomatically.TwoTypesofimagequalityassessment–Subjectivequalityassessment–Objectivequalityassessment4SubjectiveQualityMeasureThebestwaytofindqualityofanimageistolookatitbecausehumaneyesaretheultimateviewer.Subjectiveimagequalityisconcernedwithhowimageisperceivedbyaviewerandgivehisorheropiniononaparticularimage.Themeanopinionscore(MOS)hasbeenusedforsubjectivequalityassessmentfrommanyyears.Instandardsubjectivetestwherenooflistenersratetheheardaudioqualityoftestsentencesreasbybothmaleandfemalespeakeroverthecommunicationmediumbeingtested.TooInconvenient,timeconsumingandexpensive5ExampleofMOSscoreTheMOSisgeneratedbyavaragintheresultofasetofstandard,subjectivetests.MOSisanindicatoroftheperceivedimagequality.MOSscore[24]MOSscoreof1isworstimagequalityand5isbest.MeanOpinionScore(MOS)MOSQualityImpairment5ExcellentImperceptible4GoodPerceptiblebutnotannoying3FairSlightlyannoying2PoorAnnoying1BadVeryannoying6ObjectiveQualityMeasureMathematicalmodelsthatapproximateresultsofsubjectivequalityassessmentGoalofobjectiveevalutionistodevlopequantativemeasurethatcanpredictperceivedimagequalityItplaysvarietyofroles–Tomonitorandcontrolimagequalityforqualitycontrolsystems–Tobenchmarkimageprocessingsystems;–Tooptimizealgorithmsandparameters;–Tohelphomeusersbettermanagetheirdigitalphotosandevaluatetheirexpertiseinphotographing.7ObjectiveevaluationThreetypesofobjectiveevaluationItisclassifiedaccordingtotheavailabilityofanoriginalimagewithwhichdistortedimageistobecompared–Fullreference(FR)–Noreference–Blind(NR)–Reducedreference(RR)8FullreferencequalitymetricsMSEandPSNR:themostwidelyusedvideoqualitymetricsduringlast20years.SSIM:newmetric(wassuggestedin2004)showsbetterresults,thanPSNRwithreasonablecomputationalcomplexityincreasing.someothermetricswerealsosuggestedbyVQEG,privatecompaniesanduniversities,butnotsopopular.Agreatefforthasbeenmadetodevelopnewobjectivequalitymeasuresforimage/videothatincorporateperceptualqualitymeasuresbyconsideringthehumanvisualsystem(HVS)characteristics9HVS–HumanvisualsystemQualityassessment(QA)algorithmspredictvisualqualitybycomparingadistortedsignalagainstareference,typicallybymodelingthehumanvisualsystem.Theobjectiveimagequalityassessmentisbasedonwelldefinedmathematicallymodelsthatcanpredictperceivedimagequalitybetweenadistortedimageandareferenceimage.Thesemeasurementmethodsconsiderhumanvisualsystem(HVS)characteristicsinanattempttoincorporateperceptualqualitymeasures.10MSE–MeansquareerrorMSEandPSNRaredefinedas(1)(2)Wherexistheoriginalimageandyisthedistortedimage.MandNarethewidthandheightofanimage.Listhedynamicrangeofthepixelvalues.11PropertyofMSEIftheMSEdecreasetozero,thepixel-by-pixelmatchingoftheimagesbecomesperfect.IfMSEissmallenough,thiscorrespondtoahighqualitydecompressedimage.AlsoingeneralMSEvalueincreasesasthecompressionratioincreases.12Original“Einstein”imagewithdifferentdistortions,MSEvalue[6](a)OriginalImageMSE=0(b)MSE=306(c)MSE=309(d)MSE=309(e)MSE=313(f)MSE=309(g)MSE=30813SSIM–StructuralsimilarityindexRecentproposedapproachforimagequalityassessment.Methodformeasuringthesimilaritybetweentwoimages.FullreferencemetricsValueliesbetween[0,1]TheSSIMisdesignedtoimproveontraditionalmetricslikePSNRandMSE,whichhaveprovedtobeinconsistantwithhumaneyeperception.Basedonhumanvisualsystem.14SSIMmeasurementsystemFig.2.StructuralSimilarity(SSIM)MeasurementSystem[6]15ExampleimagesatdifferentqualitylevelsandtheirSSIMindexmaps[6]16EquationforSSIMIftwononnegativeimagesplacedtogetherMeanintensity(3)Standarddeviation(4)-EstimateofsignalcontrastContrastcomparisonc(x,y)-differenceofσxandσy(5)Luminancecomparison(6)C1,C2areconstant.17EquationforSSIMStructurecomparisonisconducteds(x,y)onthesenormalizedsignals(x-µx)/σxand(y-µy)/σy(7)(8)(9)(10)α,βandγareparametersusedtoadjusttherelativeimportanceofthethreecomponents.18PropertyofSSIMSymmetry:S(x,y)=S(y,x)Boundedness:S(x,y)=1Uniquemaximum:S(x,y)=1ifandonlyifx=y(indiscreterepresentationsxi=yi,foralli=1,2…….,N).19MSEvs.MSSIM20MSEvs.SSIMsimulationre