Seite - 109 - in Document Image Processing

J. Imaging 2018,4, 43 feature learningwasusedto increase therecognitionrate [29]. In thispaper,wewill conductabroader evaluationof therobustnessof themethodspreviously testedonBalinesescript,usingtheother two palmleafmanuscriptswithKhmerandSundanesescripts. In thenext sub-sections,weprovideabrief descriptionof themethods. Foradetaileddescriptionofeachmethod, interestedreaderscanrefer to ourpreviousworks. 3.3.1.HandcraftedFeatureExtractionMethods Sincethebeginningofpatternrecognitionresearch,manyfeatureextractionmethodsforcharacter recognitionhavebeenpresented in the literature. Inourpreviouswork [28],we investigatedand evaluated theperformanceof 10 feature extractionmethodswith twoclassifiers, k-NN(k-Nearest Neighbor)andSVM(SupportVectorMachine), in29different schemes forBalinesescriptonpalm leafmanuscripts.Afterevaluatingtheperformanceof those individual featureextractionmethods, wefoundthat theHistogramofGradient (HoG) featuresasdirectionalgradient-basedfeatures [9,49] (Figure5), theNeighborhoodPixelsWeights (NPW)[50] (Figure6), theKirschDirectionalEdges [50], andZoning [12,32,50,51] (Figure 7) give verypromising results. We thenproposedanew feature extractionmethodapplyingNPWonKirschedgeimages(Figure8)andconcatenatedtheNPW–Kirsch with twoother features,HoGandZoningmethod,withk-NNastheclassifier. Figure5.Therepresentationof thearrayofcells inHoG[28]. 33 3 3 3 3 23 2 2 2 3 23 1 1 2 3 3 23 1 1 2 3 23 2 2 2 3 33 3 3 3 3 Figure6.Neighborhoodpixels forNPWfeatures [28]. Figure7.TypeofZoning(fromleft toright: vertical,horizontal,block,diagonal, circular,andradial zoning) [28]. 109