Page - 41 - in Document Image Processing

J. Imaging 2018,4, 80 experimentsonhandwrittendocuments,beforeandafterslant removal. Therecognitionresults for thehandwritingofourdatabaseswerea failure,duetohavinghistoricaldocumentsor/and languagesother thanEnglish. For thePrintDBdatabase, inFigure13, thecharactererrorratevs. theartificial slantareshown,asobtainedbyacommercialOCRsystem(AdobeAcrobat). For a page slant of less than −19 or greater than 37 degrees, it is very difficult to find acorrespondencebetweenthecharactersoftheimageandtheOCRresult. Thecorrectedversionisvery wellhandled(0degrees,English).Moreover, theOCRsoftwarehandles right-slantedcharactersbetter than left-slantedcharacters. This is likelyduetoextra trainingfor italics in thecommercial software. Thecomputationalcost is less than5s forahistoricaldocument imageofsizea littlebigger than A4andresolution600dpi inacomputerwithprocessor Intel(R)Core(TM) i5-4210UCPU@1.70GHz 2.40GHz. Since the proposed systemwas built in order to help our free-segmentationword-spotting system[25],weprovideanexampleofwordspotting taskonhandwrittendocuments. It isworth mentioningthatanimprovementoftherecallofat least20%isobserved(Figure14)forthe20document imagesofGeorgeWashingtonDBand100queries. The improvementappearsextremelyhigh,which may be a result of the query being a part of the same page or collection. However, for slanted characters, theslantdegree isnotalwaysfixedandinmanycases, theslantedcharactersoverlapwith othercharacters. 0 0.2 0.4 0.6 0.8 1 Figure13.Charactererrorratevs. thedegreeofpageslantasperformedbyacommercialOCRsystem. 41