Page - 101 - in Document Image Processing

Journal of Imaging Article BenchmarkingofDocumentImageAnalysisTasksfor PalmLeafManuscripts fromSoutheastAsia MadeWinduAntaraKesiman1,2,*,DonaValy3,4, Jean-ChristopheBurie1,ErickPaulus5, MiraSuryani 5,SetiawanHadi5,MichelVerleysen2,SopheaChhun4 andJean-MarcOgier 1 1 Laboratoire InformatiqueImageInteraction(L3i),UniversitédeLaRochelle, 17042LaRochelle,France; jean-christophe.burie@univ-lr.fr (J.-C.B.); jean-marc.ogier@univ-lr.fr (J.-M.O.) 2 LaboratoryofCultural Informatics (LCI),UniversitasPendidikanGanesha,Singaraja,Bali81116, Indonesia; michel.verleysen@uclouvain.be 3 Instituteof InformationandCommunicationTechnologies,Electronic, andAppliedMathematics (ICTEAM), UniversitéCatholiquedeLouvain,1348Louvain-la-Neuve,Belgium;dona.valy@student.uclouvain.be 4 Departmentof InformationandCommunicationEngineering, InstituteofTechnologyofCambodia, PhnomPenh,Cambodia; sophea.chhun@itc.edu.kh 5 DepartmentofComputerScience,UniversitasPadjadjaran,Bandung45363, Indonesia; erick_paulus@yahoo.com(E.P.);mira.suryani@unpad.ac.id (M.S.); setiawanhadi@unpad.ac.id (S.H.) * Correspondence:made_windu_antara.kesiman@univ-lr.fr Received: 15December2017;Accepted: 18February2018;Published: 22February2018 Abstract:Thispaperpresentsacomprehensive testof theprincipal tasks indocument imageanalysis (DIA), startingwithbinarization, text linesegmentation,andisolatedcharacter/glyphrecognition, andcontinuingontowordrecognitionandtransliterationforanewandchallengingcollectionof palmleafmanuscripts fromSoutheastAsia. This researchpresentsandisperformedonacomplete datasetcollectionofSoutheastAsianpalmleafmanuscripts. It contains threedifferentscripts:Khmer script fromCambodia,andBalinesescriptandSundanesescript fromIndonesia. Thebinarization task is evaluatedonmanymethodsup to the latest in somebinarizationcompetitions. Theseam carvingmethodisevaluatedfor the text linesegmentation task, comparedtoarecentlynewtext line segmentationmethodforpalmleafmanuscripts. For the isolatedcharacter/glyphrecognitiontask, theevaluation is reportedfromthehandcraftedfeatureextractionmethod, theneuralnetworkwith unsupervisedlearningfeature,andtheConvolutionalNeuralNetwork(CNN)basedmethod. Finally, theRecurrentNeuralNetwork-LongShort-TermMemory (RNN-LSTM)basedmethod isused to analyze thewordrecognitionandtransliteration task for thepalmleafmanuscripts. Theresults from allexperimentsprovide the latestfindingsandaquantitativebenchmarkforpalmleafmanuscripts analysis for researchers in theDIAcommunity. Keywords: document imageanalysis; binarization; character recognition; text line segmentation; wordrecognition; transliteration;palmleafmanuscript;dataset;benchmark;experimental test 1. Introduction Since theworld entered the digital age in the early 20th century, the need for a document imageanalysis (DIA)systemis increasing. This isdueto thedramatic increase inefforts todigitize thevarious typesofdocument collectionsavailable, especially theancientdocumentsofhistorical relics found in various parts of theworld. Some very interesting projects on awide variety of heritage document collections can bementioned here: for example, the tranScriptorium project (http://transcriptorium.eu/) [1]; theREAD(RecognitionandEnrichmentofArchivalDocuments) project (https://read.transkribus.eu/) [2], whichworks on documents from theMiddle Ages to today,andalso focusesondifferent languages ranging fromAncientGreek tomodernEnglish; the J. Imaging 2018,4, 43 101 www.mdpi.com/journal/jimaging