Page - 107 - in Document Image Processing

J. Imaging 2018,4, 43 area, region,orwindow.Niblack’smethodproposeda local thresholdingcomputationbasedonthe localmeanandlocal standarddeviationofa rectangular localwindowforeachpixelon the image. Therectangularsliding localwindowwill cover theneighborhoodforeachpixel.Usingthisconcept, Niblack’smethodwasreportedtooutperformmanythresholding techniquesandgaveoptimal results formanydocumentcollections.However, there isstill adrawbackto thismethod. Itwas foundthat Niblack’smethodworksoptimallyonlyon the text region,but isnotwell suited for largenon-text regions of an image. The absence of text in local areas forcesNiblack’smethod todetect noise as text. Thesuitablewindowsizeshouldbechosenbasedonthecharacterandstrokesize,whichmay varyforeach image. Manyother localadaptivebinarizationtechniqueswereproposedto improve theperformanceof thebasicNiblackmethod. Forexample,Sauvola’smethodisamodifiedversion ofNiblack’smethod. Sauvola’smethodproposes a local binarization technique todealwith light texture, largevariations, anduneven illumination. The improvementoverNiblack’smethod is in theuseofadaptivecontributionofstandarddeviation indeterminingthe local thresholdonthegray values of text andnon-text pixels. Sauvola’smethodprocesses the image inN×Nadjacent and non-overlappingblocksseparately. Wolf’smethodtriedtoovercometheproblemofSauvola’smethodwhenthegrayvaluesof text andnon-textpixelsareclose toeachotherbynormalizing thecontrastandthemeangrayvalueof the image tocompute the local threshold.However,asharpchange inbackgroundgrayvaluesacross the imagedecreases theperformanceofWolf’smethod. Twoother improvements toNiblack’smethod areNICKmethodand theRaismethod. NICKmethodproposesa thresholdcomputationderived fromthebasicNiblack’smethodandtheRaismethodproposesanoptimal sizeofwindowfor the localbinarization. 3.1.3. Training-BasedBinarization Thetoptwoproposedmethods in theBinarizationChallengefor the ICFHR2016Competition ontheAnalysisofHandwrittenText in ImagesofBalinesePalmLeafManuscriptsare training-based binarizationmethods [25]. The bestmethod in this competition employs a Fully Convolutional Network(FCN). It takesacolorsubimageas inputandoutputs theprobability thateachpixel in the sub-image ispartof the foreground. TheFCNispre-trainedonnormalhandwrittendocument images with automatically generated “ground truth” binarizations (using themethodofWolf et al. [46]). TheFCNis thenfine-tunedusingDIBCOandHDIBCOcompetition imagesandtheircorresponding groundtruthbinarizations. Finally, theFCNisfine-tunedagainontheprovidedBalinesepalmleaf images. Consequently, thepixelprobabilitiesof foregroundareefficientlypredicted for thewhole imageatonceandthresholdedat0.5 tocreateabinarizedoutput image. Thesecond-bestmethoduses twoneuralnetworkclassifiers,C1 andC2, toclassifyeachpixelas backgroundornot. Twobinarizedimages,B1 andB2,aregenerated in thisstep.C1 isaroughclassifier that tries todetectall the foregroundpixels,whileprobablymakingmistakes forsomebackground pixels. C2 isanaccurateclassifier that shouldnotclassifyabackgroundpixelasa foregroundpixelbut probablymissessomeforegroundpixels. Secondly, these twobinary imagesare joinedtoget thefinal classificationresult. 3.2. TextLineSegmentation Text line segmentation is a crucial pre-processing step inmostDIApipelines. The task aims at extractingandseparating text regions into individual lines. Most line segmentationapproaches in the literature require that the input imagebebinarized. However, due to thedegradation and noise often found inhistorical documents such as palm leafmanuscripts, the binarization task is not able to produce good enough results (see Section 5.1). In this paper, we investigate two line segmentationmethods thatare independentof thebinarization task. Theseapproachesworkdirectly oncolor/grayscale images. 107