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,2012DEDICATIONTomyparen
2、ts.iiiACKNOWLEDGMENTSIamaboveallgratefultoYannLeCunandJeanPonce,forprovidingmepatientandinsightfulguidanceduringmyyearsastheirstudent.Manythanksaswelltotheothermembersofmythesiscommitteeforgivingmefeedbackandideas.IwouldliketothankFrancisBachforbeingsuchagreatinspirationandsharpdiscussant
3、;Ihavebeenveryluckytoworkwithhim,aswellasMarc’AurelioRanzato,NicolasLeRoux,KorayKavukcuoglu,andPierreSermanet,andSamyBengioandJasonWestonatGoogle.ManyideasinthisthesiswerebornwhilediscussingwithmembersoftheWillowandSierrateams,andtheComputationalandBiologicalLearningLab.Finally,Ithankmyfa
4、milyandfriendsforencouragingmeandbearingwithmeduringalltheseyears.ThisworkwassupportedbyNSFgrantEFRI/COPN-0835878toNYU,ONRcon-tractN00014-09-1-0473toNYUandbytheEuropeanResearchCouncil(VideoWorldandSierragrants).ivABSTRACTTellingcowfromsheepiseffortlessformostanimals,butrequiresmuchenginee
5、ringforcomputers.Inthisthesis,weseektoteaseoutbasicprinciplesthatunderliemanyrecentadvancesinimagerecognition.First,werecastmanymethodsintoacommonunsu-pervisedfeatureextractionframeworkbasedonanalternationofcodingsteps,whichencodetheinputbycomparingitwithacollectionofreferencepatterns,and
6、poolingsteps,whichcomputeanaggregationstatisticsummarizingthecodeswithinsomere-gionofinterestoftheimage.Withinthatframework,weconductextensivecomparativeevaluationsofmanycodingorpoolingoperatorsproposedintheliterature.Ourresultsdemonstratearobustsuperiorityofsparsecoding(whichdecomposesan
7、inputasalinearcombinationofafewvisualwords)andmaxpooling(whichsummarizesasetofinputsbytheirmaximumvalue).Wealsoproposemacrofeatures,whichimportintothepopu-larspatialpyramidframeworkthejointencodingofnearbyfeaturescommonlypracticedinneuralnetworks,andobtainsignificant
