Seite - 128 - in Document Image Processing

Journal of Imaging Article TranscriptionofSpanishHistoricalHandwritten DocumentswithDeepNeuralNetworks EmilioGranell 1,*,EdgardChammas2,LaurenceLikforman-Sulem3, Carlos-D.Martínez-Hinarejos1,ChaficMokbel 2 andBogdan-Ionut¸Cîrstea3 1 PRHLTResearchCenter,UniversitatPolitècnicadeValència,46022València,Spain; cmartine@dsic.upv.es 2 DepartmentofComputerEngineering,UniversityofBalamand,2960Balamand,Lebanon; edgard@balamand.edu.lb (E.C.); chafic.mokbel@balamand.edu.lb (C.M.) 3 InstitutMines-Télécom/TélécomParisTech,UniversitéParis-Saclay,75013Paris,France; likforman@telecom-paristech.fr (L.L.-S.);bogdan-ionut.cirstea@telecom-paristech.fr (B.-I.C.) * Correspondence: egranell@dsic.upv.es Received: 30October2017;Accepted: 2 January2018;Published: 11 January2018 Abstract: The digitization of historical handwritten document images is important for the preservation of cultural heritage.Moreover, the transcription of text images obtained from digitization isnecessary toprovideefficient informationaccess to thecontentof thesedocuments. HandwrittenTextRecognition(HTR)hasbecomeanimportant researchtopic in theareasof image andcomputational languageprocessing that allowsus toobtain transcriptions fromtext images. State-of-the-artHTRsystemsare,however, far fromperfect.Onedifficulty is that theyhavetocope with imagenoiseandhandwritingvariability.Anotherdifficulty is thepresenceofa largeamount ofOut-Of-Vocabulary(OOV)words inancienthistorical texts.Asolutionto thisproblemis touse external lexical resources,but suchresourcesmightbescarceorunavailablegiven thenatureand theageof suchdocuments. Thisworkproposes a solution toavoid this limitation. It consists of associatingapowerfulopticalrecognitionsystemthatwillcopewithimagenoiseandvariability,with a languagemodelbasedonsub-lexicalunits thatwillmodelOOVwords. Sucha languagemodeling approachreduces thesizeof the lexiconwhile increasingthe lexiconcoverage. Experimentsarefirst conductedon thepubliclyavailableRodrigodataset,whichcontains thedigitizationofanancient Spanishmanuscript, with a recognizer based onHiddenMarkovModels (HMMs). They show that sub-lexicalunitsoutperformwordunits in termsofWordErrorRate (WER),CharacterError Rate (CER) andOOVword accuracy rate. This approach is then applied to deepnet classifiers, namelyBi-directionalLong-ShortTermMemory (BLSTMs) andConvolutionalRecurrentNeural Nets (CRNNs). ResultsshowthatCRNNsoutperformHMMsandBLSTMs,reachingthe lowestWER andCERfor this imagedatasetandsignificantly improvingOOVrecognition. Keywords:historicalhandwritten transcription;out-of-vocabularywordrecognition; character-level languagemodel;wordstructureretrieval 1. Introduction Thedigitizationofhistoricalhandwrittendocument images is important for thepreservationof culturalheritage.Moreover, the transcriptionof text imagesobtainedfromdigitization isnecessary toprovideefficient informationaccess to thecontentof thesedocuments. Automatic transcription of these documents is performed by Handwriting Text Recognition (HTR) systems, which are traditionally composed of an opticalmodel, a dictionary and aLanguageModel (LM).However, HTRsystems face several challenges at both the image and languagemodeling levels. Historical document imagesmayincludedefectsduetoage,manipulationandbleed-throughof ink. Theymay also includecalligraphic initial lettersand longcharacter strokesasornaments. This isparticularly J. Imaging 2018,4, 15 128 www.mdpi.com/journal/jimaging