基于机器学习的图像质量盲评价

BlindImageQualityAssessmentBasedonMachineLearning高新波ISN国家重点实验室西安电子科技大学Nov.1~3,2013安子科技学MLA2013MLA2013OutlineƒƒImageQualityAssessment(IQA)ImageQualityAssessment(IQA)ƒƒMachineLearningandIQAMachineLearningandIQAƒƒRelatedWorkofMyGroupRelatedWorkofMyGroupƒƒOpeningIssuesOpeningIssuesVIPSLab,XidianUniversityMLA2013OutlineƒƒImageQualityAssessment(IQA)ImageQualityAssessment(IQA)ƒƒMachineLearningandIQAMachineLearningandIQAƒƒRelatedWorkofMyGroupRelatedWorkofMyGroupƒƒOpeningIssuesOpeningIssuesVIPSLab,XidianUniversityMLA2013APictureIsWorthaThousandWordsVIPSLab,XidianUniversityMLA2013ImageQualityAssessment(IQA)FidelityIntelligibilityFidelityIntelligibilityThedistortionlevelbetweenTheabilityofthetestimageonThedistortionlevelbetweenthetestandreferenceimagesintheprocessofobservationfhhTheabilityofthetestimageonprovidinginformationforhuman/machineconsumerswithrespecttthfiofthehumaneye.tothereferenceimage.VIPSLab,XidianUniversityMLA2013IQAandItsApplicationsVideoBloggingPrintingEnhancementRestorationMonitorimagequalityinqualitycontrolsystemsImageAcquisitionWatermarkingAuthenticationTele-HealthcontrolsystemsiComputerGraphicsChannelCodingIQABenchmarkimage-processingsystemsandalgorithmsRetrievalContentDeliveryCompressionBroadcastingGamingOptimizethesystemsandDeliveryDigitalQoSMonitoringsystemsandparametersettingsCameraQoSMonitoringSurveillanceVIPSLab,XidianUniversityMLA2013SubjectiveIQAVerySatisfiedVerySatisfiedRUsersatisfactionMOS4.3100VerySatisfiedVerySatisfiedSatisfiedSatisfiedSomeUsersDissatisfiedSomeUsersDissatisfied4.34.03.6908070ManyUsersDissatisfiedManyUsersDissatisfiedNearlyAllUsersDissatisfiedNearlyAllUsersDissatisfied3.12.610706050[ITU,BT.500-13]NotRecommendedNotRecommended1.00„„„„MeanOpinionScore(MOS)Themostreliableway„„yExpensiveandinefficientVIPSLab,XidianUniversityMLA2013ObjectiveIQALuminanceShift(MSE=309)OriginalImageContrastStretch(MSE=306)ImpulsiveNoise(MSE=313)JPEGCompression(MSE=309)GaussianNoise(MSE=309)Blurring(MSE=308)SpatialShift(MSE=590)„„„„Calculatepixel-wisedistances.Lackofconsiderationofhumanvisualproperty.[Wang&Bovik,IEEESPM,2009]VIPSLab,XidianUniversityImageQualityAssessment(IQA)MLA2013ImageFidelityMetricsHVSModelMetricsAccumulatephysicalerrorsoftheimageSimulatevariousaspectsoftheHVSperceptionproperty¾MeanSquaredError(MSE)¾PeakSignaltoNoiseRatio(PSNR)¾Dalyvisibledifferencespredictor(VDP)¾PerceptualDistortionMetric(PDM)SignalStructureMetricsMachineLearningMetricsIQAggDescribeimagedegradationwithperceivedchangeinStructuralinformationUtilizemachinelearningindifferentaspectsofimagequalityassessmentinformation¾StructureSimilarity(SSIM)¾Feature-Similarity(FSIM)¾BlindImageQualityIndex(BIQI)¾Blind/ReferencelessImageSpatialQUalityEvaluator(BRISQUE)VIPSLab,XidianUniversityMLA2013OutlineƒƒImageQualityAssessment(IQA)ImageQualityAssessment(IQA)ƒƒMachineLearningandIQAMachineLearningandIQAƒƒRelatedWorkofMyGroupRelatedWorkofMyGroupƒƒOpeningIssuesOpeningIssuesVIPSLab,XidianUniversityMLA2013MachineLearningMachineLearningseekstodeveloptheoriesandcomputersystemsfortiƒrepresenting;ƒclassifying,clusteringandrecognizing;ƒreasoningunderuncertainty;ƒpredicting;ƒandreactingtoƒ…complex,realworlddata,basedonthesystem'sownexperiencewithdata,and(hopefully)underaunifiedmodelormathematicalframework,thatƒcanbeformallycharacterizedandanalyzedƒcantakeintoaccounthumanpriorknowledgeƒcangeneralizeandadaptacrossdataanddomainsƒcangeneralizeandadaptacrossdataanddomainsƒcanoperateautomaticallyandautonomouslyƒandcanbeinterpretedandperceivedbyhuman.VIPSLab,XidianUniversityMachineLearningandIQAMLA2013MachineLearning:subjectiveIQAÆobjectiveIQAUtilizethemachinelearningtomodelthedifferentmodulesofthesubjectiveIQAMachineLearningforFeatureRepresentationMachineLearningforQualityPredictionMachineLearningforDistortionIdentification0.30.350.250.30.3512345678-2.5-2-1.5-1-0.500.5subbande12345678-13-12-11-10-9-8-7-6subband-50500.050.10.150.20.2512345678-2.5-2-1.5-1-0.500.5subbande12345678-13-12-11-10-9-8-7-6subband-50500.050.10.150.2ImageFeaturesFeaturesQualityFeaturesFeaturesDistortionsFeaturesVIPSLab,XidianUniversityMLA2013MachineLearningandIQAMachineLearningforFeatureRepresentation„Representimagefeaturessparselyandrecombineimagefeaturesoptimally„Obtaineffectivefeaturerepresentationsforimagequalityperception„EnhancethereliabilityofimagequalityassessmentalgorithmygqygMachineLearningRegressionDictionaryLearninggRepresentationLearningTestImagesFeaturesImageQuality…DeepLearning…VIPSLab,XidianUniversityMLA2013MachineLearningandIQALearningaBlindMeasureofPerceptualImageQualitySVRFeaturesMarginalDistribution#2230PCASVRExtractJointDistribution#14400PCATestImagesImageQualityBlur/noiseStatistics#57CompactFeaturesFeatures„ExtractfeaturesthatmeasureaspectsofimagestructureandstatisticsRddffddiPCA„RedundancyoffeaturesarereducedviaPCA„CompactfeaturesareprojectedtoqualityscoresthroughSVR„Themethodaddressesthelimitationwhichassumesthatonlyonedistortiontditithidithdldditildittittypedominatesintheimageanditcanhandleadditionaldistortiontypes[H.Tnag,N.Joshi,&A.Kapor,CVPR,2011