[CVPR 2012] Bilevel Sparse Coding for Coupled Feature Spaces

《[CVPR 2012] Bilevel Sparse Coding for Coupled Feature Spaces》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、BilevelSparseCodingforCoupledFeatureSpacesJianchaoYang†,ZhaowenWang†,ZheLin‡,XianbiaoShu†,ThomasHuang††BeckmanInstitute,UniversityofIllinoisatUrbana-Champaign,Urbana,Illinois‡AdobeSystemsInc.,SanJose,California†‡{jyang29,wang308,xshu2,huang}@illinois.edu,zlin@adobe.co

2、mAbstractofatomsfromapre-defineddictionary.Recently,therehasbeenfastgrowinginterestindictionarytraining,i.e.,usingInthispaper,weproposeabilevelsparsecodingmodelmachinelearningtechniquestolearnover-completedictio-forcoupledfeaturespaces,whereweaimtolearndictio-nariesdir

3、ectlyfromdata,sothatthemostrelevantproper-nariesforsparsemodelinginbothspaceswhileenforcingtiesofthesignalscanbeefficientlycaptured.Mostlearningsomedesiredrelationshipsbetweenthetwosignalspaces.algorithmsemploytheℓ0-orℓ1-normasthesparsitypenaltyWefirstpresentournewgener

4、alsparsecodingmodelthatmeasureforrepresentations,whichleadtosimpleoptimiza-relatessignalsfromthetwospacesbytheirsparserepresen-tionformulationsandallowtheuseofrecentdevelopedef-tationsandthecorrespondingdictionaries.Thelearningficientsparsecodingtechniques.Exampleworks

5、includealgorithmisformulatedasagenericbileveloptimizationtheMethodofOptimalDirections(MOD)withℓ0sparsityproblem,whichissolvedbyaprojectedfirst-orderstochas-measureproposedbyEnganetal.[10],thegreedyK-SVDticgradientdescentalgorithm.ThisgeneralsparsecodingalgorithmbyAharo

6、netal.[1],anefficientformulationwithmodelcanbeappliedtomanyspecificapplicationsinvolv-ℓ1sparsitymeasurebyLeeetal.[13],andanonlinedictio-ingcoupledfeaturespacesincomputervisionandsignalnarylearningalgorithmbyMairaletal.[15].Comparedprocessing.Inthiswork,wetailorourgenera

7、lmodeltowiththeconventionalmathematicallydefineddictionaries,learningdictionariesforcompressivesensingrecoveryandthelearneddictionariesaremoreadaptivetothesignaldis-singleimagesuper-resolutiontodemonstrateitseffective-tribution,whichhaveattainedstate-of-the-artperforma

8、ncesness.Inbothcases,thenewsparsecodingmodelremark-onmanysignalprocessingtasks,e.g.,denoising[1],in-ablyoutperformspreviousa