Learning a Deep Convolutional Network for image super resolution

《Learning a Deep Convolutional Network for image super resolution》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、LearningaDeepConvolutionalNetworkforImageSuper-ResolutionChaoDong1,ChenChangeLoy1,KaimingHe2,andXiaoouTang11DepartmentofInformationEngineering,TheChineseUniversityofHongKong,China2MicrosoftResearchAsia,Beijing,ChinaAbstract.Weproposeadeeplearningmethodforsingleimagesuper-resolution(SR).Ourmethoddir

2、ectlylearnsanend-to-endmappingbe-tweenthelow/high-resolutionimages.Themappingisrepresentedasadeepconvolutionalneuralnetwork(CNN)[15]thattakesthelow-resolutionimageastheinputandoutputsthehigh-resolutionone.Wefurthershowthattraditionalsparse-coding-basedSRmethodscanalsobeviewedasadeepconvolutionalnet

3、work.Butunliketraditionalmeth-odsthathandleeachcomponentseparately,ourmethodjointlyoptimizesalllayers.OurdeepCNNhasalightweightstructure,yetdemonstratesstate-of-the-artrestorationquality,andachievesfastspeedforpracticalon-lineusage.Keywords:Super-resolution,deepconvolutionalneuralnetworks.1Introduc

4、tionSingleimagesuper-resolution(SR)[11]isaclassicalproblemincomputervision.Recentstate-of-the-artmethodsforsingleimagesuper-resolutionaremostlyexample-based.Thesemethodseitherexploitinternalsimilaritiesofthesameim-age[7,10,23],orlearnmappingfunctionsfromexternallow-andhigh-resolutionexemplarpairs[2

5、,4,9,13,20,24,25,26,28].Theexternalexample-basedmethodsareoftenprovidedwithabundantsamples,butarechallengedbythedifficultiesofeffectivelyandcompactlymodelingthedata.Thesparse-coding-basedmethod[25,26]isoneoftherepresentativemeth-odsforexternalexample-basedimagesuper-resolution.Thismethodinvolvessevera

6、lstepsinitspipeline.First,overlappingpatchesaredenselyextractedfromtheimageandpre-processed(e.g.,subtractingmean).Thesepatchesarethenencodedbyalow-resolutiondictionary.Thesparsecoefficientsarepassedintoahigh-resolutiondictionaryforreconstructinghigh-resolutionpatches.Theoverlappingreconstructedpatche

7、sareaggregated(oraveraged)toproducetheoutput.PreviousSRmethodspayparticularattentiontolearningandoptimiz-ingthedictionaries[25,26]oralternativewaysofmodelingthem[4,2].However,therestofthestepsinthepipelineh