Seite - 87 - in Document Image Processing

Journal of Imaging Article HandwrittenDevanagariCharacterRecognition UsingLayer-WiseTrainingofDeepConvolutional NeuralNetworksandAdaptiveGradientMethods MaheshJangid1,*andSumitSrivastava2 1 DepartmentofComputerScienceandEngineering,SchoolofComputing&InformationTechnology, ManipalUniversity Jaipur,Rajasthan303007, India 2 Departmentof InformationTechnology,SchoolofComputing&InformationTechnology, ManipalUniversity Jaipur,Rajasthan303007, India; sumit.310879@gmail.com * Correspondence:mahesh_seelak@yahoo.co.in;Tel.:+91-0141-3999279 Received: 6December2017;Accepted: 12February2018;Published: 13February2018 Abstract:Handwrittencharacter recognition iscurrentlygetting theattentionof researchersbecause ofpossibleapplications inassisting technologyforblindandvisually impairedusers,human–robot interaction,automaticdataentry forbusinessdocuments, etc. In thiswork,weproposea technique torecognizehandwrittenDevanagari charactersusingdeepconvolutionalneuralnetworks (DCNN) whichareoneof therecent techniquesadoptedfromthedeeplearningcommunity.Weexperimented theISIDCHARdatabaseprovidedby(InformationSharingIndex) ISI,KolkataandV2DMDCHAR databasewithsixdifferentarchitecturesofDCNNtoevaluate theperformanceandalso investigate theuse of six recentlydeveloped adaptive gradientmethods. A layer-wise technique ofDCNN hasbeen employed that helped to achieve thehighest recognition accuracy andalsoget a faster convergencerate. Theresultsof layer-wise-trainedDCNNarefavorable incomparisonwith those achievedbyashallowtechniqueofhandcraftedfeaturesandstandardDCNN. Keywords:handwrittencharacter recognition;deep learning;Devanagari characters; convolutional neuralnetwork;adaptivegradientmethods 1. Introduction In the last fewyears,deeplearningapproaches[1]havebeensuccessfullyappliedtovariousareas suchas imageclassification, speech recognition, cancer celldetection, video search, facedetection, satellite imagery, recognizing trafficsignsandpedestriandetection,etc. Theoutcomeofdeep learning approaches is alsoprominent, and in somecases the results are superior tohumanexperts [2,3] in thepastyears.Mostof theproblemsarealsobeingre-experimentedwithdeeplearningapproaches with theview toachieving improvements in the existingfindings. Different architectures ofdeep learninghavebeen introduced in recent years, suchasdeep convolutional neural networks, deep beliefnetworks, andrecurrentneuralnetworks. Theentire architecturehas showntheproficiency in different areas. Character recognition is one of the areaswheremachine learning techniques havebeenextensivelyexperimented. Thefirstdeep learningapproach,which isoneof the leading machine learningtechniques,wasproposedforcharacter recognition in1998onMNISTdatabase [4]. Thedeep learning techniques are basically composedofmultiple hidden layers, andeachhidden layerconsistsofmultipleneurons,whichcompute thesuitableweights for thedeepnetwork.Alotof computingpower isneededtocompute theseweights, andapowerful systemwasneeded,which wasnoteasilyavailableat that time. Since then, theresearchershavedrawntheirattentiontofinding thetechniquewhichneeds lesspowerbyconvertingthe images into featurevectors. In the last few decades,a lotof featureextractiontechniqueshavebeenproposedsuchasHOG(histogramoforiented J. Imaging 2018,4, 41 87 www.mdpi.com/journal/jimaging