Natural Image Denoising with自然图像去噪 卷积网络

《Natural Image Denoising with自然图像去噪 卷积网络》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、NaturalImageDenoisingwithConvolutionalNetworksVirenJain1H.SebastianSeung1;21Brain&CognitiveSciences2HowardHughesMedicalInstituteMassachusettsInstituteofTechnologyMassachusettsInstituteofTechnologyAbstractWepresentanapproachtolow-levelvisionthatcombinestwomainideas:theuseofconvolutionalnetworksasan

2、imageprocessingarchitectureandanunsu-pervisedlearningprocedurethatsynthesizestrainingsamplesfromspecificnoisemodels.Wedemonstratethisapproachonthechallengingproblemofnaturalimagedenoising.Usingatestsetwithahundrednaturalimages,wefindthatcon-volutionalnetworksprovidecomparableandinsomecasessuperiorpe

3、rformancetostateoftheartwaveletandMarkovrandomfield(MRF)methods.Moreover,wefindthataconvolutionalnetworkofferssimilarperformanceintheblindde-noisingsettingascomparedtoothertechniquesinthenon-blindsetting.WealsoshowhowconvolutionalnetworksaremathematicallyrelatedtoMRFapproachesbypresentingameanfieldth

4、eoryforanMRFspeciallydesignedforimagedenois-ing.Althoughtheseapproachesarerelated,convolutionalnetworksavoidcompu-tationaldifficultiesinMRFapproachesthatarisefromprobabilisticlearningandinference.Thismakesitpossibletolearnimageprocessingarchitecturesthathaveahighdegreeofrepresentationalpower(wetrai

5、nmodelswithover15,000param-eters),butwhosecomputationalexpenseissignificantlylessthanthatassociatedwithinferenceinMRFapproacheswithevenhundredsofparameters.1BackgroundLow-levelimageprocessingtasksincludeedgedetection,interpolation,anddeconvolution.Thesetasksareusefulbothinthemselves,andasafront-end

6、forhigh-levelvisualtaskslikeobjectrecog-nition.Thispaperfocusesonthetaskofdenoising,definedastherecoveryofanunderlyingimagefromanobservationthathasbeensubjectedtoGaussiannoise.Oneapproachtoimagedenoisingistotransformanimagefrompixelintensitiesintoanotherrep-resentationwherestatisticalregularitiesar

7、emoreeasilycaptured.Forexample,theGaussianscalemixture(GSM)modelintroducedbyPortillaandcolleaguesisbasedonamultiscalewaveletde-compositionthatprovidesaneffectivedescriptionoflocalimagestatistics[1,2].Anotherappro