资源描述:
《2012 Learning Hierarchical Feature Extractors For Image Recognition .pdf》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、LearningHierarchicalFeatureExtractorsForImageRecognitionbyY-LanBoureauAdissertationsubmittedinpartialfulfillmentoftherequirementsforthedegreeofDoctorofPhilosophyDepartmentofComputerScienceNewYorkUniversitySeptember2012YannLeCunJeanPoncecY-LanBoureauAllRightsReserved,2012DEDICATIONTomyparents.i
2、iiACKNOWLEDGMENTSIamaboveallgratefultoYannLeCunandJeanPonce,forprovidingmepatientandinsightfulguidanceduringmyyearsastheirstudent.Manythanksaswelltotheothermembersofmythesiscommitteeforgivingmefeedbackandideas.IwouldliketothankFrancisBachforbeingsuchagreatinspirationandsharpdiscussant;Ihavebe
3、enveryluckytoworkwithhim,aswellasMarc’AurelioRanzato,NicolasLeRoux,KorayKavukcuoglu,andPierreSermanet,andSamyBengioandJasonWestonatGoogle.ManyideasinthisthesiswerebornwhilediscussingwithmembersoftheWillowandSierrateams,andtheComputationalandBiologicalLearningLab.Finally,Ithankmyfamilyandfrien
4、dsforencouragingmeandbearingwithmeduringalltheseyears.ThisworkwassupportedbyNSFgrantEFRI/COPN-0835878toNYU,ONRcon-tractN00014-09-1-0473toNYUandbytheEuropeanResearchCouncil(VideoWorldandSierragrants).ivABSTRACTTellingcowfromsheepiseffortlessformostanimals,butrequiresmuchengineeringforcomputers
5、.Inthisthesis,weseektoteaseoutbasicprinciplesthatunderliemanyrecentadvancesinimagerecognition.First,werecastmanymethodsintoacommonunsu-pervisedfeatureextractionframeworkbasedonanalternationofcodingsteps,whichencodetheinputbycomparingitwithacollectionofreferencepatterns,andpoolingsteps,whichco
6、mputeanaggregationstatisticsummarizingthecodeswithinsomere-gionofinterestoftheimage.Withinthatframework,weconductextensivecomparativeevaluationsofmanycodingorpoolingoperatorsproposedintheliterature.Ourresultsdemonstratearobustsuperiorityofsparsecoding(whichdecomposesaninputasalinearcombinatio
7、nofafewvisualwords)andmaxpooling(whichsummarizesasetofinputsbytheirmaximumvalue).Wealsoproposemacrofeatures,whichimportintothepopu-larspatialpyramidframeworkthejointencodingofnearbyfeaturescommonlypracticedinneuralnetworks,andobtainsignificant