Seite - 122 - in Document Image Processing

J. Imaging 2018,4, 43 recognizinganddirectly transliteratingBalinesewords. For theKhmerandSundanesedatasets, the LSTMarchitectureseemstostruggle to learn the trainingdata.Moresyntheticdata trainingwitha more frequentwordshouldbegenerated inorder tosupport the trainingprocess. For theBalinese dataset, asequencedepthof100pixelswithaneuronsizeof200givesabetter result forbothLSTM andBLTSMarchitecture.Mostof theSoutheastAsianscriptsaresyllabic scripts.Onecharacter/glyph in these scripts represents a syllable,with a sequence of letters inLatin script. In this case,word transliteration isnot justwordrecognitionwithone-to-oneglyph-to-letterassociation. Thismakes wordtransliterationmorechallengingthancharacter/glyphrecognition. Table8.Experimental results forwordrecognitionandtransliterationtasks (in%errorrate for test). Methods(withOCRopy[56]Framework) Balinese Khmer Sundanese BLSTM1(seq_depth60,neuronsize100) 43.13 Latin text: 73.76Khmer text: 77.88 75.52 LSTM1(seq_depth100,neuronsize100) 42.88 - - BLSTM2(seq_depth100,neuronsize200) 40.54 - - LSTM2(seq_depth100,neuronsize200) 39.70 - - Figure26.Errorrate forBalinesewordrecognitionandtransliterationtest set. Figure27.Errorrate forKhmerwordrecognitionandtransliterationtest set. 122