Seite - 112 - in Document Image Processing

J. Imaging 2018,4, 43 architecturewith two-directional (forward andbackward) context processing. LSTMarchitecture iswidely evaluatedas ageneric and language-independent text recognizer [55]. In thiswork, the OCRopy(https://github.com/tmbdev/ocropy) [56] framework isusedto testandevaluate theword recognitionandtransliteration tasks for thepalmleafmanuscriptcollection.OCRopyprovides the functional library of theOCRsystembyusingRNN-LSTMarchitecture (http://graal.hypotheses. org/786) [57,58].Weevaluated thedatasetwithunidirectionalLSTMandthe (BidirectionalLTSM) BLSTMarchitecture. 4. Experiments:DatasetsandEvaluationMethods Fromthe threemanuscriptcorpuses (Khmer,Balinese,andSundanese), thedatasets foreachDIA taskwereextractedandusedin theexperimentalworkfor this research. 4.1. Binarization 4.1.1.Datasets Thepalm leafmanuscript datasets for binarization task arepresented inTable 1. ForKhmer manuscripts, one ground truth binarized image is provided for each image, but for Balinese and Sundanesemanuscripts,eachimagehastwodifferentgroundtruthbinarizedimages[17,25]. Thestudy ofgroundtruthvariabilityandsubjectivitywasreported in thepreviouswork[24]. In this research, we only used the first binarized ground truth image for evaluation. The binarized ground truth images forKhmermanuscriptsweregeneratedmanuallywith thehelpofphoto editing software (Figure11).Apressure-sensitive tipstylus isusedto traceeach text strokebykeepingtheoriginal size of thestrokewidth[59]. For themanuscripts fromBali, thebinarizedgroundtruth imageshavebeen createdwithasemi-automaticscheme[17,23–25] (Figure12). Thebinarizedgroundtruth images for Sundanesemanuscriptsweremanually [22]generatedusingPixLabeler [60] (Figure13). The training set isprovidedonlyfor theBalinesedataset.Weusedall imagesof theKhmerandSundanesecorpuses asa test setbecause the training-basedbinarizationmethod(ICFHRG1method,seeSection5.1)was evaluatedfor theKhmerandSundanesedatasetsbyusingonly thepre-trainedBalinese trainingset weightedmodel. Table1.Palmleafmanuscriptdatasets forbinarizationtask. Manuscripts Train Test GroundTruth Dataset Balinese 50pages 50pages 2×100pages ExtractedfromAMADI_LontarSet [17,25,40] Khmer - 46pages 1×46pages ExtractedfromEFEO[20,59] Sundanese - 61pages 2×61pages ExtractedfromSundaDataset ICDAR2017[22] Figure11.Khmermanuscriptwithbinarizedgroundtruth image. 112