Page - 93 - in Document Image Processing

J. Imaging 2018,4, 41 Figure2.Layer-wise trainingofdeepconvolutionalneuralnetwork. Algorithm1.Layerwise trainingofdeepconvolutionalneuralnetwork INPUT:Model,T, t,α1,α2,n\\T=(TrainData), t= (TestData), OUTPUT:TM \\TrainedModel Begin\\Addfirst layerofconvolutional layerandpooling layer Model.add (xCy,T,Relu) Model.add (xPy) Model.add (xFC) Model.add (xOU) Model.compile (optimizer) Model.fit (T, t,α1) forall I :=1: n-1step1do \\Removethe last twolayers (FC&OU) ofexistingmodel toaddnext layerofconvolutionalandpooling Model.layer.pop() Model.layer.pop() Model.add (xCy,T,Relu) Model.add (xPy) \\Againaddedfullyconnectedandoutput layer Model.add (xFC) Model.add (xOU) Model.compile (optimizer) Model.fit (T, t,α1)\\Trainedthemodelwithhigh learningrate endfor Model.fit (T, t,α2)\\Performfinetuningwith lowlearningrate end 4. ExperimentsandDiscussions Experimentswerecarriedoutontwodatabases: ISIDCHARandV2DMDCHARusingtheDCNN, layer-wiseDCNNanddifferentadaptivegradientmethods.As it ishardtodelineate thenumberof layersofDCNNthatcanproduce thebest result,weconsideredsixdifferentnetworkarchitectures (NA) ofDCNNas shown inTable 1. NA-1 contains only single convolutional-pooling layer and 500 fully connected neurons to observe the first response ofDCNN. The next, NA-2 has double thenumberof fullyconnectedneurons. Theaimis toobserve the impactofenhancement. Further, NA-3andNA-4have twoC-P layerswithvariation in thenumberofkernels toanalysis the impactof twoC-Players. The last,NA-5andNA-6havethreeC-P layers. Initially, thedifferentnetworkarchitecturesofDCNNwereappliedoneachdatabase tofindout thebestmodel for thatparticulardatabaseandthentheproposedlayer-wiseDCNNwasappliedto observe the impactof thatmodel. Themodelshavealsobeentestedwithdifferentadaptivegradient methodsto thesemethods; theyarealsounderexperiment toobserve theirperformance.Ourwork alsoshowsthe impactofdifferentadaptivegradientmethodsonrecognitionaccuracy. 93