Yi-Ma (super-resolution)【深度学习】.ppt

《Yi-Ma (super-resolution)【深度学习】.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、ImageSuper-resolutionviaSparseRepresentationJianchaoYang,JohnWright,YiMaCoordinatedScienceLaboratoryDepartmentofElectricalandComputerEngineeringUniversityofIllinoisatUrbana-ChampaignSIAMImagingScienceSymposium,July7,2008Super-resolutionassparserepresentationindictionaryof

2、rawimagepatchesSolutionvia-normminimizationGlobalconsistency,featureselectionExperiments:QualitativecomparisontopreviousmethodsQuantitativecomparisontopreviousmethodsConclusionsandDiscussionsOUTLINELEARNING-BASEDSUPER-RESOLUTION–ProblemformulationProblem:givenasinglelow-r

3、esolutioninput,andasetofpairs(high-andlow-resolution)oftrainingpatchessampledfromsimilarimages,reconstructahigh-resolutionversionoftheinput.OutputOriginalInputTrainingpatchesAdvantage:morewidelyapplicablethanreconstructive(manyimage)approaches.Difficulty:single-imagesuper

4、-resolutionisanextremelyill-posedproblem.LEARNING-BASEDSUPER-RESOLUTION–PriorworkHowshouldweregularizethesuper-resolutionproblem?Markovrandomfield[Freemanet.Al.IJCV‘00]Primalsketchprior[Sunet.Al.CVPR‘03]Neighborembedding[Changet.Al.CVPR‘04]Softedgeprior[Daiet.Al.ICCV‘07]?

5、LEARNING-BASEDSUPER-RESOLUTION–PriorworkHowshouldweregularizethesuper-resolutionproblem?Markovrandomfield[Freemanet.Al.IJCV‘00]Primalsketchprior[Sunet.Al.CVPR‘03]Neighborembedding[Changet.Al.CVPR‘04]Softedgeprior[Daiet.Al.ICCV‘07]?Ourapproach:High-resolutionpatcheshaveasp

6、arselinearrepresentationwithrespecttoanovercompletedictionaryofpatchesrandomlysampledfromsimilarimages.outputhigh-resolutionpatchhigh-resolutiondictionaryforsomewithLINEARSPARSEREPRESENTATION–SRasCompressedSensingWedonotdirectlyobservethehighresolutionpatch,butrather(feat

7、uresof)itslow-resolutionversion:Theinputlow-resolutionpatchsatisfieslinearmeasurementsofsparsecoefficientvector!dictionaryoflow-resolutionpatches.downsampling/blurringoperatorLINEARSPARSEREPRESENTATION–SRasCompressedSensingIfwecanrecoverthesparsesolutiontotheunderdetermin

8、edsystemoflinearequations,wecanreconstructasconvexrelaxationThisproblemcanbeefficientlysolvedbyl