[ICPR 2012 Andrew] End-to-End Text Recognition with Convolutional Neural Networks

《[ICPR 2012 Andrew] End-to-End Text Recognition with Convolutional Neural Networks》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、End-to-EndTextRecognitionwithConvolutionalNeuralNetworksTaoWang∗DavidJ.Wu∗AdamCoatesAndrewY.NgStanfordUniversity,353SerraMall,Stanford,CA94305{twangcat,dwu4,acoates,ang}@cs.stanford.eduAbstract32×3225×25×965×5×964×4×2562×2×256Fullend-to-endtextrecogn

2、itioninnaturalimages[Non-Text]isachallengingproblemthathasreceivedmuchatten-tionrecently.Traditionalsystemsinthisareahavere-[Text]liedonelaboratemodelsincorporatingcarefullyhand-ConvolutionConvolutionClassificationengineeredfeaturesorlargeamountsofpr

3、iorknowl-AveragePoolingAveragePoolingedge.Inthispaper,wetakeadifferentrouteandcom-binetherepresentationalpoweroflarge,multilayerFigure1.CNNusedfortextdetection.neuralnetworkstogetherwithrecentdevelopmentsinunsupervisedfeaturelearning,whichallowsustou

4、searelatedfieldssuchasvisualrecognition[3]andactioncommonframeworktotrainhighly-accuratetextdetec-recognition[7].Inthecaseoftextrecognition,thetorandcharacterrecognizermodules.Then,usingonlysystemin[2]achievescompetitiveresultsinbothtextsimpleoff-the-

5、shelfmethods,weintegratethesetwodetectionandcharacterrecognitionusingasimpleandmodulesintoafullend-to-end,lexicon-driven,scenescalablefeaturelearningarchitectureincorporatingverytextrecognitionsystemthatachievesstate-of-the-artlittlehand-engineeringa

6、ndpriorknowledge.performanceonstandardbenchmarks,namelyStreetWeintegratetheselearnedfeaturesintoalarge,ViewTextandICDAR2003.discriminatively-trainedconvolutionalneuralnetwork(CNN).CNNshaveenjoyedmanysuccessesinsimi-larproblemssuchashandwritingrecogni

7、tion[8],visual1Introductionobjectrecognition[1],andcharacterrecognition[16].Byleveragingtherepresentationalpowerofthesenet-Extractingtextualinformationfromnaturalimagesworks,weareabletotrainhighlyaccuratetextdetectionisachallengingproblemwithmanyprac

8、ticalapplica-andcharacterrecognitionmodules.Usingthesemod-tions.Unlikecharacterrecognitionforscanneddocu-ules,wecanbuildanend-to-endsystemwithonlysim-ments,recognizingtextinunconstrainedimagesiscom-plepost-processingtechniqueslikenon-maximalsup-plica