Seite - 111 - in Document Image Processing

J. Imaging 2018,4, 43 Thenetworkconsistsof threesetsofconvolutionandmaxpoolingpairs.All convolutional layersuse a strideofoneandarezeropaddedso that theoutput is the samesizeas the input. Theoutputof eachconvolutional layer isactivatedusingtheReLufunctionandfollowedbyamaxpoolingof2×2 blocks. Thenumbersof featuremaps (of size5×5)used in the threeconsecutiveconvolutional layers are8, 16, and32, respectively. Theoutputof the last layers isflattened,andafully-connected layer with1024neurons (alsoactivatedwithReLu) is added, followedby the last output layer (softmax activation)consistingofNclassneurons,whereNclass is thenumberofcharacterclasses.Dropoutwith probabilityp=0.5 isappliedbefore theoutput layer topreventoverfitting.Wetrainedthenetwork usinganAdamoptimizerwithabatchsizeof100anda learningrateof0.0001. Figure10.Architectureof theCNN. 3.4.WordRecognitionandTransliteration Inorder tomakethepalmleafmanuscriptsmoreaccessible, readable,andunderstandable toa wideraudience,anoptical character recognition(OCR)systemshouldbedeveloped. InmanyDIA systems,wordor text recognition is thefinal task in theprocessingpipeline.However,normally in SoutheastAsianscript thespeechsoundof thesyllablechange is relatedtosomecertainphonological rules. In thiscase,anOCRsystemisnotenough. Therefore,a transliterationsystemshouldalsobe developedtohelp transliterate theancientscriptsonthesemanuscripts. Bydefinition, transliteration is defined as the process of obtaining the phonetic translation of names across languages [54]. Transliteration involves renderinga language fromonewritingsystemtoanother. In [54], theproblem isstatedformallyasasequencelabelingproblemfromonelanguagealphabet toanother. Itwillhelpus to indexandtoquicklyandefficientlyaccess thecontentof themanuscripts. Inourpreviouswork[29], acompleteschemeforsegmentation-basedglyphrecognitionandtransliterationspecific toBalinese palmleafmanuscriptswasproposed. In thiswork,asegmentation-freemethodwillbeevaluatedto recognizeandtransliterate thewords fromthreedifferentscriptsofapalmleafmanuscript. RNN/LSTM-BasedMethods Fromthe lastdecade, sequence-analysis-basedmethodsusingaRecurrentNeuralNetwork-Long Short-TermMemory(RNN-LSTM)typeoflearningnetworkhavebeenverypopularamongresearchers in text recognition. RNN-LSTM-basedmethodtogetherwithaConnectionistTemporalClassification (CTC)worksasasegmentation-free learning-basedmethodtorecognize thesequenceofcharacters in aword or textwithout any handcrafted feature extractionmethod. The raw image pixel can be sent directly as the input to the learning network and there is no requirement to segment the trainingdatasequence. RNNisbasicallyanextendedversionof thebasic feedforwardneuralnetwork. In a RNN, the neurons in the hidden layer are connected to each other. RNNoffers very good context-awareprocessing torecognizepatterns inasequenceor timeseries.OnedrawbackofRNNis thevanishinggradientproblem. Todealwith thisproblem, theLSTMarchitecturewas introduced. TheLSTMnetworkaddsmultiplicativegatesandadditive feedback. BidirectionalLSTMisanLSTM 111