Page - 129 - in Document Image Processing

J. Imaging 2018,4, 15 true forSpanishdocuments fromthe16thcenturyasseen inFigure1.Ancient textsalso includerare characters,grammatical forms,wordspellingsandnamedentitiesdistinct frommodernones. Such forms lead toOut-Of-Vocabulary (OOV)words, i.e.,words thatdonotbelong to thedictionaryof theHTRsystem. ImprovingHTRsystemsatboth imageand language levels isan important issue for the recognitionof suchancient historical documents. Themaingoal of this paper is todesign efficientHTRsystemsthatprocessdocument imageswritten inSpanishandthatcancopewithancient character formsandlanguage. Figure1.Sample imageofaSpanishdocument fromthe16thcentury. Several approacheshavebeenproposed tobuildopticalmodels for handwriting recognition. Suchapproaches includeHiddenMarkovModels (HMMs) [1–4],RecurrentNeuralNetworks (RNNs) suchasLongShort-TermMemory(LSTMs)andtheirvariants: Bi-directionalLSTMs(BLSTMs)and Multi-DimensionalLSTMs(MDLSTMs) [5].HMMsenableembeddedtrainingandcanberobust to noiseandlineardistortions.However,RNNsandtheirvariantsaregenerativemodels thatperform better thanHMMsintermsofaccuracy.Nowadays,RNNscanbetrainedbyusingdedicatedresources such asGraphic ProcessorUnits (GPUs) that considerably reduce training time. ByusingGPUs, RNNscanbe trained inasimilar amountof timerequired to trainHMMswith traditionalCentral ProcessingUnits (CPUs). Usually, the inputsofHMMsandRNNsaresequencesofhandcraftedfeaturesorpixel columns. However, deep learning approaches starting with convolutional layers as the first layers allow extracting learning-basedfeatures insteadofhandcraftedones [6–8]. Generally, inHTRsystems, theopticalmodelsareassociatedwithdictionaries (lexicalmodels) andLanguageModels (LMs), usually at theword level, in order to direct the recognition of real words andplausibleword sequences (see Figure 2). In order to build open vocabulary systems, languagemodelsbasedoncharacterunitscanbeused[9]. Then, thedictionary is limitedto theset 129