Seite - 121 - in Document Image Processing

J. Imaging 2018,4, 43 Figure23.Text linesegmentationofBalinesemanuscriptwith theSeamCarvingmethod(green)and AdaptivePathFinding(red). Figure24.Text linesegmentationofKhmermanuscriptwiththeSeamCarvingmethod(green)and AdaptivePathFinding(red). Figure25.Text linesegmentationofSundanesemanuscriptwith theSeamCarvingmethod(green) andAdaptivePathFinding(red). 5.3. IsolatedCharacter/GlyphRecognition Theexperimental results for isolatedcharacter/glyphrecognitiontaskarepresentedinTable7. Forhandcraftedfeaturewithk-NN, theKhmersetwith113,206 train imagesand90,669 test images will needaconsiderableamountof time forone-to-onek-NNcomparison, sowedonot think it is reasonabletouseit. ForCNN1,previousworkonlyreportedresults for theBalineseset. Forall ICFHR competitionmethods, thecompetitionwasproposedonly for theBalineseset, soweonlyhave the reported results for theBalinese set. According to these results, thehandcrafted featureextraction combinationofHoG-NPW-Kirsch-Zoning isaproperchoiceresulting inagoodrecognitionrate for BalineseandKhmer characters/glyphs. TheCNNmethodsalso showsatisfactory results, but the differences inrecognitionratesarenot toosignificantwith thehandcraftedfeaturecombinations. The unbalancednumberof imagesamplesforeachcharacterclassmeanstheCNNmethoddidnotperform optimally. For theSundanesedataset, thehandcrafted featurewithNNslightlyoutperformed the CNNmethod. TheUFLmethodslightly increasedtherecognitionrateof thepureNNmethodfor the KhmerandBalinesedatasets. Table7.Experimental results for isolatedcharacter/glyphrecognitiontasks (in%recognitionrate). Methods Balinese Khmer Sundanese HandcraftedFeature (HoG-NPW-Kirsch-Zoning)withk-NN[28] 85.16 - 72.91 HandcraftedFeature (HoG-NPW-Kirsch-Zoning)withNN[29] 85.51 92.15 79.69 HandcraftedFeature (HoG-NPW-Kirsch-Zoning)withUFL+NN[29] 85.63 92.44 79.33 CNN1[28] 84.31 - - CNN2 85.39 93.96 79.05 ICFHRG1:VCMF[25] 87.44 - - ICFHRG1:VMQDF[25] 88.39 - - ICFHRG3[25] 77.83 - - ICFHRG5[25] 77.70 - - 5.4.WordRecognitionandTransliteration Theexperimental results forwordrecognitionandtransliterationtaskarepresented inTable8. Theerrorrates forwordrecognitionandtransliterationtestssetoneachtrainingmodel iterationare showninFigures26–28. TheLSTM-basedarchitectureofOCRopyseemsverypromising in termsof 121