Seite - 130 - in Document Image Processing

J. Imaging 2018,4, 15 of different characters, and the transitionprobabilities between the charactermodels aregivenby a character LM.Character-basedLMsare alsouseful for related tasks such asword spotting [10]. In thepreviouscharacterLMapproachoreveningeneralwordLMapproaches, theopticalmodels stillmodelcharacters.However, inworkssuchas [11,12], theopticalmodelsmodelstrokes thatare concatenatedto formwords. OpticalModel LanguageModel y pequeños LexicalModel y p e q u e ñ o s ypequeños Figure2.Schemeofahandwritten text recognitionsystem. Whenaword-baseddictionaryhelps therecognitionprocess, thehandwritingrecognitionsystem canonly transcribea limitednumberofwords. Thesizeof thedictionary isacompromisebetween atoo largesizeyieldingwordconfusionsandatoosmalloneyieldingmanyunknownwords.Words of the test set that arenotpresent in theHTRdictionaryaredenotedasOut-Of-Vocabulary (OOV) words. Several typesofOOVwordsexist, suchascommonwordsusinga lesscommongrammatical form,misspellings,wordsattachedtopunctuationmarks,hyphenatedwordsorwordscontainingrare characters (abbreviations, special signs,etc.). AnapproachtocopewithOOVwordsconsistsofextendingthedictionarywithexternal lexical resources, such asWikipedia [13], or in the case of historical documents, with the transcription of other documents from the same period and topic [14]. From these resources, the language model can also be refined. However, in the general case, such resourcesmay not be available, andaproportionofwords (suchasnamedentities andrarewords) still remainsasOOV.Another approach for copingwithOOVwordsconsistsofmodeling textat a sub-word level, asa sequence ofcharacters, syllablesormulti-grams[15].Hybridapproaches [16,17]consistofusingword-based languagemodels for themost frequentwordsandcharacter-basedmodels for the less frequentones. Insub-wordapproaches, thedictionary isconsiderablyreducedto thenumberof lexicalunits, aswell as the computational complexity. In addition, the languagemodel canmodelunknownwordsby combiningsuch lexicalunits. In thiswork,we compare severalHTR systems, based onHMMs, RNNs and convolutional RNNs(CRNNs). TheCRNNis inspiredfromaverydeeparchitecturepresented in [18]. It consists of stackingBLSTMsandassociating themwithconvolutional layers. Featuresare thusautomatically extractedbytheconvolutional layersandprocessedbytheBLSTMlayers.Wealsomodeldictionaries and language models of our HTR systems with sub-word units. We apply this approach to therecognitionofapubliclyavailableSpanishhistoricaldocumentsdataset.WecompareseveralHTR systemsbasedondifferenttypesofsub-wordunits,andweshowthatsub-wordunitsaremoreefficient thanwordunits.Weobtain, toourknowledge, thebest recognitionresultsonthisSpanishdatasetby associatingsub-wordunitswith thedeepestHTRoptical system,namelytheCRNN.Wealsoobtain highrates for therecognitionofOOVwords. The rest of the paper is structured as follows: the Spanish historical manuscript used in theexperimentation ispresented in thenextsection(Section2); theHTRsystemsandtheexperimental conditionsaredescribedinSection3;ourexperimentsandtheobtainedresultsarereportedinSection4; theconclusionsandfutureworkaredrawninSection5;finally, inAppendixA,several recognition examplesareshown. 2.TheRodrigoDataset TheRodrigo corpus [19]was obtained from thedigitization of the book “Historia de España del arçobispoDonRodrigo”,written in ancient Spanish in 1545. It is a singlewriter bookwhere mostpages consist of a single blockofwell-separated lines of calligraphical text, as the examples 130