Page - 29 - in Document Image Processing

Journal of Imaging Article SlantRemovalTechniqueforHistorical DocumentImages ErginaKavallieratou1,*,LaurenceLikforman-Sulem2andNikosVasilopoulos1 1 Department InformationandCommunicationSystemsEngineering,Universityof theAegean,Samos83200, Greece;nvasilopoulos@aegean.gr 2 InstitutMines-Télécom/TélécomParisTech,UniversitéParis-Saclay,75013Paris,France; laurence.likforman@telecom-paristech.fr * Correspondence: kavallieratou@aegean.gr Received: 14May2018;Accepted: 5 June2018;Published: 12 June2018 Abstract:Slantedtexthasbeendemonstratedtobeasalient featureofhandwriting. Itsestimation is anecessarypreprocessing task inmanydocument imageprocessingsystems inorder to improve therequired training. Thispaperdescribesandevaluatesanewtechnique for removing theslant fromhistoricaldocumentpages thatavoids thesegmentationprocedure into text linesandwords. Theproposedtechniquefirst reliesonslantangledetectionfromanaccurateselectionof fragments. Then,aslant removal technique isapplied.However, thepresentedslant removal techniquemay becombinedwithanyotherslantdetectionalgorithm.Experimental resultsareprovidedfor four document imagedatabases: twohistoricaldocumentdatabases, theTrigraphSlantdatabase (theonly databasededicatedtoslant removal), andaprinteddatabase inorder tochecktheprecisionof the proposedtechnique. Keywords: slant removal;document imageprocessing;document imagepage 1. Introduction Inhandwriting, slant removal isanecessarycomponentof the textnormalizationprocedure in systemsthatperformrecognition(e.g.,optical character recognition(OCR) [1]orword-spotting[2]), in order to improvethe trainingprocedure (lesssamples, lowercomputationalcost).Moreover,writer identification/verificationsystemsalsouse slant estimationand/ordetection [3]. After ideal slant removalprocessing, the text shouldappearwith thevertical stokesparallel to theperpendicularaxis of thepage. Due to its importance,manyresearchershavealreadydeveloped techniques for slant removal [4–17]. Theavailable techniquesmaybedividedinto threecategories: 1. Techniques thatestimate theslantbyaveragingtheanglesof thenear-vertical strokes [4–7]. 2. Techniques thatanalyzeprojectionhistograms[8,9]anddetect theslantbasedonapre-defined criterion(e.g., aparametermaximizationorminimization). 3. Techniques thatarebasedonthestatisticsofchain-codedcontours [10–12]. Consideringtheapplicationthese techniquescanhandle, theycanbefurtherclassifiedinto: 1. Uniformslantestimationandremoval techniques [4–12]: theydealwithuniformslantallover the text. 2. Non-uniformslantcorrection techniques [13–16]: theyhandle thecharactersapartanddealwith theexistenceofseveral slants, simultaneously. J. Imaging 2018,4, 80 29 www.mdpi.com/journal/jimaging