Seite - 98 - in Document Image Processing

J. Imaging 2018,4, 41 We found that theNA-6networkarchitecturewithRMSPropoptimizerproduced thehighest recognitionaccuracy. Thisnetworkwasagain trainedbylayer-wisemodelasdescribed inSection3.3. Thisnetworkwas testedwith ISIDCHAR,V2DMDCHAR,andcombineddatabases. Theresults are reported in Table 6. It has been seen that a nice enhancement in the recognition accuracy was recorded by the layer-wise trainingmodel. The 97.30% recognition accuracywas obtained on ISIDCHAR database and 97.65% recognition accuracy obtained on V2DMDCHAR database. The layer-wise trainingmodelwasalsoappliedaftercombiningboth thedatabasesandobtained98% recognitionaccuracywhen70%of thesampleswereusedfor trainingandtherestused for testing. Thecurrentwork iscomparedtopreviousworksonISIDCHARdatabase inTable7. Table6. Inthis table,wereportedthemaximumrecognitionaccuracyobtainedwithNA-6andRMSProp optimizeronISIDCHAR,V2DMDCHARandcombinedbothwhenthemodelwastrainedlayer-wise. Database No. ofSamples RecognitionAccuracy DCNN Layer-WiseDCNN ISIDCHAR 36,172 96.02% 97.30% V2DMDCHAR 20,305 96.45% 97.65% ISIDCHAR+V2DMDCHAR 56,477 96.53% 98.00% Table7.Comparisonof recognitionaccuracybyotherresearchers. S.No. AccuracyObtained Feature;Classifier MethodProposedby DataSize 1 95.19 Gradient;MIL U.Pal [26] 36,172 2 95.24 GLAC;SVM M.Jangid[32] 36,172 3 96.58 Masking,SVM M.Jangid[33] 36,172 4 96.45 DCNN Proposedwork 36,172 5 97.65 SL-DCNN Proposedwork 36,172 6 98 SL-DCNN Proposedwork 56,477 5.Conclusions Deeplearning isoneof theprominent technologies thathavebeenexperimentallystudiedwith entiremajor areas of computer vision and document analysis. In this paper, we experimentally developedadeepconvolutionalneuralnetwork(DCNN)andadaptivegradientmethods torecognize the unconstrained handwritten Devanagari characters. The deep convolutional neural network helped us to find the best features automatically and also classify them. We experimentedwith a handwrittenDevanagari character databasewith six differentDCNNnetwork architectures as wellassixdifferentoptimizers. Thehighest recognitionaccuracy96.02%wasobtainedusingNA-6 networkarchitectureandRMSProp—anadaptivegradientmethod (optimizer). Further,weagain trainedDCNNlayer-wise,which is also adoptedbymany researchers to enhance the recognition accuracy,usingNA-6networkarchitectureandtheRMSPropadaptivegradientmethod.UsingDCNN layer-wise trainingmodel, our database obtained 98% recognition accuracy,which is the highest recognitionaccuracyof thedatabase. Acknowledgments:Theauthorsarethankful to theISI,KolkatatoprovideadatabaseandtheManipalUniversity Jaipur toprovide thesupercomputing facilitywithout this facilitydeep learningconceptwasnotpossible for thehandwrittenDevanagari characters. AuthorContributions:MaheshJangidhasenvisagedthestudy,designedtheexperiments,andwrotethemanuscript. SumitSrivastavaperformedsomepartofexperimentsandcorrectedthemanuscript. Bothauthorsreadandapproved thefinalmanuscript. Conflictsof Interest:Theauthorsdeclare that theyhavenocompeting interests. 98