Page - 1 - in Document Image Processing

Journal of Imaging Editorial DocumentImageProcessing LaurenceLikforman-Sulem1,*andErginaKavallieratou2 1 InstitutMines-Télécom/TélécomParisTech,UniversitéParis-Saclay,75013Paris,France 2 Department InformationandCommunicationSystemsEngineering,Universityof theAegean,Samos83200, Greece;kavallieratou@aegean.gr * Correspondence: laurence.likforman@telecom-paristech.fror likforman@telecom-paristech.fr Received: 15 June2018;Accepted: 15 June2018;Published: 22 June2018 Keywords: document image processing; preprocessing; document restoration; binarization; slant removal; text-line segmentation; handwriting recognition; indic/arabic/asian scripts;OCR; Video OCR; word spotting; retrieval; document datasets; performance evaluation; document annotationtools The Special Issue “Document ImageProcessing” in the Journal of Imaging aims at presenting approacheswhichcontribute toaccess thecontentofdocument images. Theseapproachesarerelated to lowlevel taskssuchas imagepreprocessing, skew/slantcorrections,binarizationanddocument segmentation, aswell as high level tasks such as OCR, handwriting recognition, word spotting or script identification. This special issuebrings together 12papers that discuss suchapproaches. Thefirst threearticlesdealwithhistoricaldocumentpreprocessing. TheworkbyHanifetal. [1]aims at removingbleed-throughusinganon-linearmodel, andat reconstructing thebackgroundbyan inpaintingapproachbasedonnon-localpatchsimilarity. ThepaperbyAlmeidaetal. [2]proposesa newbinarizationapproachthat includesadecision-basedprocess forfindingthebest thresholdfor eachRGBchannel. In thepaperbyKavallieratouetal. [3], a segmentation-freeapproachbasedonthe Wigner-Villedistribution isusedtodetect theslantofadocumentandcorrect it. Onceadocument image ispreprocessed,anextstepdescribed in thepaperbyGhoshetal. [4] consists inseparatingtextcomponents fromnon-textones,usingaclassifierbasedonLBPfeatures. Followingstepsmayconsist inrecognizingtextcomponentsorsearchingfromwordqueries. In the paperbyNashwanet al. [5] aholistic-basedapproach for the recognitionofprintedArabicwords isproposed, coupledwithanefficientdictionaryreduction. In theworkbyNagendaretal. [6] it is shownthatusingaqueryspecific fastDynamicTimeWarpingdistance, improves theDirectQuery Classifier (DQC)wordspottingsystem. Deepneuralnetwork-basedapproachesarenowwidelyusedin thedomainofdocument image processing,especially for therecognitionof textualelements. Thefollowingpapersalso followthis trend. In theworkbyJangidandSrivastava[7],deepconvolutionalnetworks trainedlayer-wise,are appliedto therecognitionofDevanagari characters. ThepaperbyKesimanetal. [8] isdedicatedto southeastAsianscriptswrittenonpalmleafs.CharacterandwordimagesarerecognizedbyCNNs (ConvolutionalNeuralNetworks) andRNNs (RecurrentNeuralNetworks), respectively. Several binarizationandtext-linesegmentationapproachesarealsobenchmarkedonthesespecificdocuments. TheworkbyGranelletal. [9]describesanefficient text-linerecognitionsystem,basedonCNNand stacksofRNNs, thathasbeendevelopedfor therecognitionofhistoricalSpanishdocuments. These documents includeout-of-vocabularyancientwordswhicharehandledbya languagemodelbasedon sub-lexicalunits. Annotateddatasets are necessary to train systemsor to evaluate the various tasks related to document image processing. In several papers published in this special issue, newdatasets are releasedaswell asopen-source tools that areable togenerate synthetic images. Adatasetof indic scripts is releasedin thepaperbyMukhopadhyayetal. [10]andfirst resultsareprovidedwith this J. Imaging 2018,4, 84 1 www.mdpi.com/journal/jimaging