2009 10th International Conference on Document Analysis and Recognitio - PDF document

825 826 827 2009 10th International Conference on Document Analysis and Recognition978-0-7695-3725-2/09 $25.00 © 2009 IEEEDOI 10.1109/ICDAR.2009.176823 824 ItalicorRoman:WordStyleRecognitionWithoutAPrioriKnowledgeforOld PrintedDocuments LorisEynard,HubertEmptoz Universit ´ edeLyon,CNRS INSA-Lyon,LIRIS,UMR5205, F-69621,France f loris.eynard,hubert.emptoz g @insa-lyon.fr Abstract ThispaperpresentsanItalic/Romanwordtyperecog- nitionsystemwithoutaprioriknowledgeonthecharac- ters'font.Thismethodaimsatanalyzingolddocumentsin whichcharactersegmentationisnottrivial.Thereforeour approachsegmentsthedocumentintowordsandanalyse thetextwordperword.Todenethewordstyle,wecom- binethreecriteriawhicharebasedonthevisualdifferences betweenawordandaslantedversionofthesameword. Thesecriteriaaredenedthankstofeaturescomputedfrom theverticalprojectionproleoftheword.Becausewedo notassumeaspecicslantangle,wecomputethesemea- suresonawholerangeofpossibleslantanglesandthen sumtheobtainedscores.Ourresultsshowaratioof100% recognitionforItalicwordsand97.2%forRomanwords. 1.Introduction OurworkwasdesignedtondItalicwordsinhistorical newspaperofthe18 th century,morespecicallyintheim- agesofthe GazettedeLeyde dataset. Commercialsoftwareforcharacterrecognitionarenotef- cientforrecognizingthecharacterinolddocuments.On the GazettedeLeyde datasetweobtainaratioof88%of characterrecognition 1 when99%isconsideredtobeef- cient.Ofcourseonwellconservedpartofthedocuments theresultsareverygoodbutondamagedpartsthisratiode- creasesharply.TheItalicwords,duetothethicknessofthe Italiccharacter,aremoresensitivetothetimedamagesthan theRomanwordandsotheyarebadlyrecognizedbycom- mercialOCR. ThoseItalicwordsrepresentparticularnouns(patronymic orlocationnames)andsoareveryinterestingforre- searchersinHumanSciences.Thenalaimistoextract 1 withABBYYFineReader8Pro wordstypesetinItalicinparagraphstypesetinRoman inseveralyearsofGazettewhichrepresentsthousandsof pages. Figure1.Italicstylenounsinthemiddleofa Romanstyleparagraph Becauseofthislargeamountofdocuments,ourmethod mustbeabletodecideverypreciselyifawordiseitherin ItalicorRomanstyle.Thedifcultiesweencounteredare duetothetheconservationoftheolddocumentsandtheir digitization,whichresultsinseveraltypesofdegradations, e.ginkbleed-through,holes,inkfading,etc.These artefactscreatelinksbetweencharacters(andevenmore betweenItaliccharacterswhicharecloserthanRoman ones)whichmakesthemhardertosegment.Forthisreason weproposeacharacters'segmentation-freemethod.We baseourselfonthevisualcharacteristicsofthecharacters ofawordthatcanbeinterpretedbyanalysingthevertical projectionproleofthewordimage.Thoseanalysisbased onthreevisualfeatureswouldgiveusscorestodecide whetherawordisinItalicorRomanstyle.Aswedonot supposeaspecicslantangleoftheItalictype,whichmay varysignicantlyacrossthedocumentweareprocessing, wetestarangeofanglesratherthanonlyconsideringone slantangle. Thispaperisorganizedasfollows:rstwerecallthestate oftheartonthisproblem.Thenwedescribeourmethod andconcludewiththeresults. left)seemsvisuallycorrect. Inthenextthreesectionswedescribethethreecriteriawe usetodenetheItalicstyle.Theyarebasedonthediffer- encesbetweenawordinItalicstyleandthesamewordin Romanstyle. 3.2.Firstcriterion:Verticalblackcolumn ThemaindifferencebetweenanItalicandRomanver- sionofawordisthepresenceoflongverticalstrokesinthe caseoftheRomanversion,whicharerepresentedaspeaks intheverticalprojectionprole. Figure3.highestblackcolumnofhistogram, seethatmaximumofhistogramdiffersfrom originalandslantedimages Slantingaverticalstrokeintheimagetranslatesintoat- teningthecorrespondingpeakintheprojectionprole(see gure3).ThedifferencebetweenaRomanstylewordand anItalicstyleversionofitisgivenbycomparingthemax- imaoftheirrespectiveprojectionproles. Weobtainedtwovalues MH o fortheoriginalwordand MH s fortheslantedone.Therstcriterion,called C 1 is valuedat1if MH s �MH o and0otherwise. 3.3.SecondCriterion:Overlappingofthe characters ConsideringanItalicstyleword,weobservethatthetop ofaslantedcharacteroftenoverlapthebottomofthefol- lowing.Thisobservationwillbethesameforaninverted- Italicstyleword. Evenifthereisnotarealoverlappingthenthewhite spacebetweenthetwocharactersisnoticeablyreduced. Thiscriterionshowsthisdifferencebetweenthewhite spacesintheRomanversionandintheItalicversionofa word(seeg.4). Thisfeatureistranslatedintheverticalprojectionpro- lebyanalysingthewhitespacebetweenblacksections. Eachblacksectionrepresentsacharacter(ormoreifthere Figure4.Overlappingofcharactersin slantedword(whitespacearemarkedby darkverticallinesfortheslantedwordand bylightverticallinesfortheoriginalword) isoverlapping)andthewhitespacerepresentsthespacebe- tweencharacters.Intheprojectionprole,wesearchallthe whitespacesbetweentwoblackpixels.Bydeductionthe morewhite-between-blackpixelthereis,themoretheword isRomanstyle.Let'scall W o thetotalwidthofwhitespace intheoriginalimageand W s intheslantedword.Wegive thevalue1tothesecondcriterion C 2 if W o �W s and0 otherwise. 3.4.Thirdcriterion:Variationoftheslops oftheVerticalHistogram Thelastcriterionthatweusedisthevariationofslopein theverticalprojectionproleoftheword.Foreachword, thecharactersarerepresentedaspeaksofblackpixelsin theverticalprojectionprole.IfthiswordisRomanstyle, thenthosepeakswouldhavevaluesofslopecertainlyhigher thanforaslantedword.Moreovertheseslopeswillvary muchmoresuddenlythanforanItalicword.Thiscould beexplainedbecauseanItalicstylecharacterappearsmore spreadintheverticalprojectionprolethanawellstraight Romanone.Thesedifferencesofslopetranslatethemore horizontalconcentrationoftheblackpixelforanon-slanted character. Tocomputethevariationofslopeoftheprojectionpro- lewecomputethesecondderivativeoftheverticalprojec- tionprole.Asfortherstcriterionitwouldmakenosense toconsideronlythemaximumofthesevariations.Thatis whywecomputeanaverageofthetenmaximumvariations ofslopeoftheprojectionprole.Ifweconsider VS o theav- eragemaximavariationofslopefortheoriginalwordand VS s fortheslantedimage.Thenthiscriterion, C 3 getthe value1if VS o �VS s and0ifnot. 3.5.FinalStyleDecision Weobtainthreebinarycriteria C 1 , C 2 and C 3 giving usindicationsonthestyleoftheword.Consideringonly thosebinarycriteriawillgiveusatooarbitrarydecision fortheword'sstyle.Tospecifythisdecision,accordingto thegroundtruth,wedeneaweightforeachcriterion.By combiningthecriterionanditsassociatedweightweobtain ascoretodecidetheword'sstyle.Theseweightsrepre- sentaratioofwordsverifyingthecriteriaaccordingtotheir style.Forexample w ro 1 representthepercentageofRoman stylewordsverifyingtherstcriterion.Thesepercentage arecomputedononly40wordsofeachstyle.Thecom- putedweightvaluesareshownintable1. Style Italic Roman criteria C 1 0.15 0.8 C 2 1 0.5 C 3 1 0.4 Table1.Weightsvalues Thankstothecriteriaandtheirassociatedweightswe candenedecisiontermsforeachstyle.Wecomputethese decisiontermsforeachangle  asfollow: T Roman  = 3 X i =1 ( w ro i :C i +(1 � w ro i ) : (1 � C i )) T Italic  = 3 X i =1 ( w it i :C i +(1 � w it i ) : (1 � C i )) . T Roman  istheRomanstyledecisiontermforawordfor theskewangle  . T Italic  istheItalicstyledecisiontermof thesamewordforthesameslantangle  .Wecomparethe valuesofthesetermstodecideifthewordisItalicorRoman styleforthespecicslantangle  .If T Roman  �T Italic  thenthewordischaracterizedasRomanstylefortheslant angle  andviceversa.Accordingtothis,wecall D  the decisionfortheangle  describedasfollow: D  =  RomanifT Roman  �T Italic  Italicifnot Inolddocuments,theslantangleoftheItalicstyleisnot xedlikeinrecent'sone.Wecannotassumeaspecic skewanglebutwesupposeItalicstyletobeslantedbetween 5and20degrees.Wetestalltheseanglesbeforechoosing theword'sstyle.Weassumethatifawordischaracterized Romanstyleinmostofthe15anglesthatwetest,thenit wouldbexedasRomanstyle.Wecall D thenaldecision foraword'sstyle.Todenethisdecision D weadaptthe KroneckerDeltanoted  s;D  to:  s;D  =  1 ifs = D  0 ifnot Thenaldecision D givesusthestyleforwhichthesum foreachangleofthisadaptatedKroneckerDeltaismaxi- mized.Thenthenaldecision D isdenedasfollow: D =argmax s = Roman;Italic 20 X  =5  s;D  Ifwecallintermediatedecisionfortheangle  the D  . Theresultofthefunction D isthestylewhichgivesthe maximumofintermediatedecisionsbysummingthedeci- sionforallfteenpossibleslantangles. D resultinatwo choicesdecisionforthewordtobeeitherRomanorItalic style. 4.ResultsandDiscussion OurapproachwasdesignedtodetectItalicstylewords suchaspropernounsinthe GazetteofLeyde dataset.Those nounsaresupposedtobepatronymicsortoponymswhich aremostlylargewordsof4lettersormore.Forthisreason wedon'texpectgoodresultsonshortwords(wordscon- taininglessthan3letters).Moreover,ourmethodtakesinto accountoverlappingcharacters.Foratwoletterswordthere isonlyonepossibleoverlappingandtwoforathreeletters word.Thisobservationdecreasesignicantlytheinuence ofthecriterion C 2 onthenaldecision.Table2showsthe word'slengthhistogramforthedatasetof1358wordswe usedinourexperiments. wordssize(inletters) 23  3 Total Italic 1514176 205 Roman 269145739 1153 Table2.Numberoftestingwords Thisdistributionnaturallyarosefromthedatasetandhas notbeeninuencedbyus.ThesmallnumberofItalicwords isrelatedtothespecictypesettingofthedocuments. wordssize(inletters) 23  3 total Italic 100100100 100 Roman 89.597.299.99 97.2 Table3.RecognitionRates(%) Weobtainreallygoodresultsfordecidingwhethera wordisItalicstyleorRomanstyle.Asexpected,ourresults arelowerforshortwords(2or3letters)butstillverygood. Asthismethodischaractersegmentation-free,itcanrecog- nizeword'sstylesonoldorblurreddocuments,aswellas documentscontainingwordswithtouchingcharactersWe expectthismethodtoworkforanyclassofdocumentsand withanycharacterstyle. 4.1.Acknowledgments Thisworktakepartofa ClusterCulture,Patrimoineet Cr ´ eation whichisaregionalresearchclusteroftheRegion Rhone-Alpes,France. References [1]B.B.ChaudhuriandU.Garain.Automaticdetectionofitalic, boldandall-capitalwordsindocumentimages.In ICPR '98:Proceedingsofthe14thInternationalConferenceon PatternRecognition-Volume1 ,page610,Washington,DC, USA,1998.IEEEComputerSociety. [2]K.-C.FanandC.H.Huang.Italicdetectionandrectication. J.Inf.Sci.Eng. ,23(2):403–419,2007. [3]E.Kavallieratou,N.Fakotakis,andG.Kokkinakis.Slant estimationalgorithmforocrsystem. PatternRecognition , 34:2515–2522,2001. [4]Y.Li,S.Naoi,M.Cheriet,andC.Y.Suen.Asegmentation methodfortouchingitaliccharacters.In ICPR'04:Proceed- ingsofthePatternRecognition,17thInternationalConfer- enceon(ICPR'04)Volume2 ,pages594–597,Washington, DC,USA,2004.IEEEComputerSociety. [5]H.MaandD.Doermann.Adaptivewordstyleclassication usingagaussianmixturemodel.In ICPR'04:Proceedings ofthePatternRecognition,17thInternationalConferenceon (ICPR'04)Volume2 ,pages606–609,Washington,DC,USA, 2004.IEEEComputerSociety. [6]E.K.PavlosStathisandN.Papamarkos.Anevaluationsurvey ofbinarizationalgorithmsonhistoricaldocuments. ICPR'08: Proceedingsofthe19thInternationalConferenceonPattern Recognition . [7]C.SunandD.Si.Skewandslantcorrectionfordocumentim- agesusinggradientdirection.In ICDAR'97:Proceedingsof the4thInternationalConferenceonDocumentAnalysisand Recognition ,pages142–146,Washington,DC,USA,1997. IEEEComputerSociety. [8]L.Zhang,Y.Lu,andC.L.Tan.Italicfontrecognitionusing strokepatternanalysisonwaveletdecomposedwordimages. In ICPR'04:ProceedingsofthePatternRecognition,17th InternationalConferenceon(ICPR'04)Volume4 ,pages835– 838,Washington,DC,USA,2004.IEEEComputerSociety.