2013.Image super-resolution using local learnable kernel regression

《2013.Image super-resolution using local learnable kernel regression》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、ImageSuper-ResolutionUsingLocalLearnableKernelRegressionRenjieLiaoandZengchangQinIntelligentComputingandMachineLearningLabSchoolofAutomationScienceandElectricalEngineeringBeihangUniversity,Beijing,Chinalrjconan@gmail.com,zcqin@buaa.edu.cnAbstract.Inthispaper,weaddresstheproble

2、moflearning-basedim-agesuper-resolutionandproposeanovelapproachcalledLocalLearnableKernelRegression(LLKR).Theproposedmodelemploysalocalmet-riclearningmethodtoimprovethekernelregressionforreconstructinghighresolutionimages.WeformulatethelearningproblemasseekingmultipleoptimalMa

3、halanobismetricstominimizethetotalkernelre-gressionerrorsonthetrainingimages.Throughlearninglocalmetricsinthespaceoflowresolutionimagepatches,ourmethodiscapabletobuildaprecisedata-adaptivekernelregressionmodelinthespaceofhighresolutionpatches.Sincethelocalmetricssplitthewholed

4、atasetintoseveralsubspacesandthetrainingprocesscanbeexecutedoff-line,ourmethodisveryefficientatruntime.Wedemonstratethatthenewdevel-opedmethodiscomparableorevenoutperformsothersuper-resolutionalgorithmsonbenchmarktestimages.Theexperimentalresultsalsoshowthatouralgorithmcanstillac

5、hieveagoodperformanceevenwithalargemagnificationfactor.1IntroductionThebasicideaofSuper-Resolutionistoestimateahighresolution(HR)imagefromasingleorseveraloriginallowresolution(LR)images.Itisaninherentlyill-posedinverseproblemsincethemappingbetweenHRimageandLRimageismany-to-onea

6、ndmuchinformationislostintheHR-to-LRprocess.Variousmethodshavebeenproposedtosolvethisunderdeterminedmapping.Roughly,theycanbedividedintothreemajorcategories:(1)InterpolationbasedmethodsthatgenerateHRimageusingsingleLRimage[6,7].(2)ReconstructionbasedmethodsusingmultipleLRimage

7、swithsomesmoothnesspriors[8,9].And(3)learningbasedmethods(orexamplebasedmethods)thatusealargetrainingsetofHR/LRpatchpairs[10,19].Thoughtheimplementationofinterpolationbasedmethodsareveryfast,theyareunabletoproducesharpedgesandcleardetails.Asforconventionalreconstructionbasedme

8、thods,sincetheaccuratesubpixelmotionestimationisextremelydiffic