Seite - 95 - in Document Image Processing

J. Imaging 2018,4, 41 accuracyby increasing thenumberofkernelsofconvolutional layer. ThenetworkarchitecturesNA-2, NA-4andNA-6hadmorekernels thanNA-1,NA-3andNA-5andtheyproducedhigherrecognition accuracyasobservedinTable2. Thenumberof trainableparameters foreachnetworkarchitecture is showninTable3. Theentirenetworkarchitecturewasalso testedusing theRMSPropoptimizer, andtheresultshavereported inTable4. TheNA-6networkproduced96.02%recognitionaccuracy withRMSPropwhile95.58%withAdam.ThebehaviorofNA-6withRMSPropateachepochcanbe seen inFigure4. (SRFK 1$ 1$ 1$ 1$ 1$ 1$ Figure3. Inthisfigure,wedrawtherecognitionaccuracyobtainedwithdifferentnetworkarchitectures onISIDCHARdatabaseateachepoch. TheAdamoptimizerwasused. Table 2. In this table,we report the results in termofmaximum,minimum,mean, and standard deviation recognitionaccuracyobtainedwithdifferentnetworkarchitectureson ISIDCHARwhen thesystemtrainedfor50epochswith theAdamoptimizer. Thebest scoresare inbold. RecognitionAccuracy DifferentNetworkArchitectures NA-1 NA-2 NA-3 NA-4 NA-5 NA-6 Maximum 0.8571 0.8654 0.9153 0.9224 0.9324 0.9558 Minimum 0.7208 0.7701 0.8237 0.8363 0.8077 0.8385 Average 0.8436 0.8549 0.9000 0.9058 0.9190 0.9427 Std.Deviation 0.0204 0.0169 0.0165 0.0158 0.0178 0.0168 Table3.Listof trainableparameters ineachnetworkarchitecture. NetworkArchitectures LayerType LayerSize TrainableParameters TotalParameters NA-1 Conv1layer 64×64×64 1088 34,873,135Dense layer 500 34,848,500 Output layer 47 23,547 NA-2 Conv1layer 64×64×64 1088 61,553,135Dense layer 1000 61,505,000 Output layer 47 47,047 NA-3 Conv1layer 32×64×64 544 7,265,007 Conv2 layer 32×33×33 16,416 Dense layer 1000 7,201,000 Output layer 47 47,047 95