Page - 30 - in Document Image Processing

J. Imaging 2018,4, 80 Recently, Brink et al. [3] categorized the proposed techniques by angle-frequency and repeated-shearingapproaches thataredescribedas follows: 1. Angle-frequencyapproach:Down-strokesarefirst locatedbasedonsuchcriteriaas theminimum verticalextentorvelocity.Next, theangleof the local inkdirection ismeasuredat these locations andtheresultinganglesareagglomerated inahistogram. Fromthishistogram, theslantangle is determined. This isaone-stepprocedure. 2. Repeated-shearing approach: Thismethod is basedon the assumption that theprojectionof dark pixels ismaximized along an axis parallel to the slant angle. The basic principle is to repeatedly shear imagesof individual text lines, varying the shear angle, andoptimizing the verticalprojectionofdarkpixels. Thisapproach isclearlymore timeconsuming,butprovesmore accurate,as indicatedbyitspopularity. Thefirst categorywill be referred to here as ‘slant estimation’ (one-stepprocedure), and the secondcategory is referredtoasslantdetection, since thismethodsearchesamongmany, for themost commonangle. Slantestimationtechniquesarepresented in [4–7],whereasaslantdetection technique ispresentedin[9].AccordingtoBrinketal. [3], theslantdetectiontechniquesare themostpopular with themostprecise results. The techniquedescribed in [9] isalsoused in thatpaperwhereextensive experimentsoverslantareperformed.Lastbutnot least, in thespecificexperiments, thepageswere shearedentirely, since thealternative lineorwordsegmentation ischaracterizedas“less reliableand breaks inktracesat regionboundaries” [3]. Theproposedtechniquesuptonowrequire lineorword segmentation inorder tobeapplied. InFigure1,anexampleof theslant removalalgorithmdescribed in [9], ispresented. The image is fromtheIAMHandwritingDatabase (IAM-DB),andtheapplication of thealgorithmrequires imagesegmentation into text lines (Figure1,horizontal stripes). For this example, text linesegmentationcouldsucceedsince text linesarespacedenough. It isnot thecase for thedocument image shown inFigure2 (17th century)which includes touchingascenders and descendersandnoise in the inter-linespace. Sinceallexistingalgorithmsperformslantremovalon wordor text line level, asegmentation-freeapproach isdesirable fordifficult tosegmentdocuments. Moreover,avoidingthe text-linesegmentationprocessing iscomputationally lessexpensive. Figure1.Anexampleofaslant removalapplicationresultingfromthedetectionalgorithmdescribed in [9]. Text-linesegmentation is requiredprior toslantestimation. Apreliminaryapproachhasbeendescribedin[17],while in thispaper theparametersetupis consideredanddescribed indetail.Moreover, theapproach isextensivelyevaluatedonnewdatabases. Theproposedtechnique isappropriate forslantdetectionandremoval fromdocument imageswith 30