杨志政专业英语.pptVIP

ATechniquetoPreserveEdgeInformationinSingleImageSuperResolution学生：杨志政导师：王春兴

Theproposednovelapproachpottedtheinformationaroundcurvatureedgesandhardedges(abrupttransitioninintensity)usingNonSub-SampledContourletTransform(NSCT)basedlearningprocess.Thesmoothnessofsmoothedges(gradualtransitioninintensity)ispreservedbyusingsoftedgesmoothnesspriorasaregularizingparameter.Abstract

ThreeSteps1FundamentalsofNSCTThemultiscaledecompositionpropertyofNSCTdecomposesanimageintodifferentfrequencybandsnamelylowfrequencybandandhighfrequencyband

ThreeSteps2RobustestimationofNSCTcoefficientsconstructthetrainingdatasetproperlyestimateHRcoefficientsTheinversetransformofthisimageresults

ThreeSteps3GlobalconstraintandremovalofedgeartifactsThesmoothedgeprior|I|gcanbedefinedas:Thecostfunctioncanberepresentedas:FinalHRimageisobtainedbyminimizing“E”throughanInteriveBackProjectionalgorithm.

ResultsandDiscussion

ResultsandDiscussion

Perceptioncomparisonshowsminimumedgeartifactsfortheproposedapproach.TheproposedapproachoffersdesiredresultslikeimprovedPSNRandhigherFSIMvalue.FortestimageI,PSNRisimprovedby10.6dBwithrespecttoBicubicinterpolationbecausethisimagecontainscurveedges.

ResultsandDiscussion

Theproposedalgorithmproducesmoreresults(1)

ResultsandDiscussionTheproposedalgorithmproducesmoreresults(2)

TheproposedlearningbasedalgorithmisbasedonthefactthateverycoefficientatthecoarserleveloftheNSCTcanberelatedtothecoefficientatthenextfinerlevelforthesimilarorientation.TheprobableoutliersinlearningprocessarerejectedbyadoptingrobustLorentzianerrornorm.Theimageobtainedthroughthislearningprocesspreservesinformationalongt