Deep Neural Networks – Applications in Handwriting Recognition英文学习资料

《Deep Neural Networks – Applications in Handwriting Recognition英文学习资料》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、DeepNeuralNetworks–ApplicationsinHandwritingRecognitionThéodoreBluchetheodore.bluche@gmail.comSãoPauloMeetup-9Mar.20172WhoamI?ThéodoreBluchePhDdefendedatUniversitéParis-Sudin2015DeepNeuralNetworksforLargeVocabularyHandwrittenTextRecognitionNowworkingasaReasearchEngineerata

2、2iainParis…automaticdocumentprocessing(handwritingrecognitionandmore…)…partoftheresearchteam(6people)…implementationofnewneuralnetworks…improvingthespeedandaccuracyofproductionmodels…buildthemodelsoftomorrow3HandwritingRecognition…Goal:Convertscanneddocument(image)totext4…isfullofchallenges5PuzzleWh

3、atcharactersarethose?6Answer:u,(partof)m,en,n!→hardtosegmentcharacters,thenrecognize!OfflineHandwritingRecognition“Look!”hewentonearnestly.“You’renot…towriteupPericlesinsomewayorother…you?”“Whatistheretowriteup?”saidBawley.“Thereareforty-threesubmarines..theNavy-whyshouldIpickonPericles?”Johneyedhim

4、steadily.“It’sthePrincipalbusiness,isn’tit?”hesaidquietly.Line(orWord)SegmentationTextRecognitionLanguageModel➔Challenges○theinputisavariable-sizedtwo-dimensionalimage○theoutputisavariable-sizedsequenceofcharacters○thecursivenatureofhandwritingmakesapriorsegmentationintocharactersdifficult8Outlineof

5、thistalk➔StandardHandwritingRecognition(HWR)System◆Imageprocessing-FeatureExtraction-OpticalModel-HiddenMarkovModel-LanguageModel➔DeepNeuralNetworksforHWR--pluggingNNsinthesystem◆NeuralNets:Multilayerperceptrons

6、RecurrentNeuralNetworks◆DeepNeuralNets…◆…automaticallylearngoodfeaturesandcontext➔End-to

7、-EndHWR--frompixelstotext◆MultidimensionalRecurrentNeuralNetworks◆Attention-basedmethods9Outlineofthistalk➔StandardHandwritingRecognition(HWR)System◆Imageprocessing-FeatureExtraction-OpticalModel-HiddenMarkovModel-LanguageModel➔DeepNeuralNetworksforHWR--pluggingNNsinthesystem◆NeuralNets:Multilayerpe

8、rceptrons

9、RecurrentNeuralNetworks◆DeepNeuralNets…◆…automaticallylearngoodfeaturesandcontext➔End-to-EndHWR--frompixelstotext◆MultidimensionalRecurrentNeuralNetworks◆Attention-basedmethods➔TipsandTricks