Seite - 108 - in Document Image Processing

J. Imaging 2018,4, 43 3.2.1. SeamCarvingMethod ArvanitopoulosandSüsstrunk[47]proposedabinarization-freemethodbasedonatwo-stage process:medial seamandseparatingseamcomputation. Theapproachcomputesmedial seamsby splittingthe inputpage image intocolumnswhosesmoothedprojectionprofilesare thencalculated. Thepositionsof themedial seamsareobtainedbasedonthe localmaxima locationsof theprofiles. Thegoalof thesecondstageof theapproach is tocomputeseparatingseamswith theapplicationon theenergymapwithin thearearestrictedbythemedial seamsof twoneighboring lines foundinthe previousstage. The techniquecarvespaths that traverse the imagefromleft to right, accumulating energy. Thepathwith theminimumcumulativeenergy is thenchosen. 3.2.2.AdaptivePathFindingMethod ThisapproachwasproposedbyValyet al. [27]. Themethod takesas inputagrayscale image of adocumentpage. Connected components are extracted from the input imageusing the stroke width informationbyapplyingthestrokewidth transform(SWT)ontheCannyedgemap. Thesetof extractedcomponents (filteredtoremovecomponents thatcomefromnoiseandartifacts) isusedto createastrokemap.Usingcolumn-wiseprojectionprofilesontheoutputmap,estimatednumberand medialpositionsof text linecanbedefined. Toadaptbetter toskewandfluctuation,anunsupervised learningcalledcompetitive learning isappliedonthesetofconnectedcomponents foundpreviously. Finally,apathfindingtechnique isapplied inorder tocreateseambordersbetweenadjacent linesby usingacombinationof twocost functions: onepenalizingthepath thatgoes throughthe foreground text (intensitydifferencecost functionD) andanotherone favoring thepath that stays close to the estimatedmedial lines (vertical distance cost functionV). Figure 4 illustrates an example of an optimalpath. Figure4.Anexampleofanoptimalpathgoingfromstart stateS1 togoal stateSn. 3.3. IsolatedCharacter/GlyphRecognition In aDIA system,word or text recognition tasks are generally categorized into twodifferent approaches: segmentation-basedandsegmentation-freemethods. Insegmentation-basedmethods, the isolatedcharacter recognition task isavery importantprocess [9].Aproper featureextractionand acorrectclassifierselectioncan increase therecognitionrate [48].Althoughmanymethodsfor isolated characterrecognitionhavebeendevelopedandtested,especially forLatin-basedscriptsandalphabets, there is still aneedfor in-depthevaluationof thosemethodsasappliedtovariousotherscripts. This includes the isolatedcharacterrecognitiontaskformanySoutheastAsianscripts,andmorespecifically scripts thatwerewrittenonancientpalmleafmanuscripts. Previousstudiesonisolatedcharacter recognition inpalmleafmanuscriptshavealreadybeen reported, butonlywith theBalinese script as thebenchmarkdataset [28,29]. In thatfirstwork, an experimental studyon feature extractionmethods for character recognitionofBalinese scriptwas performed[28]. For thesecondwork,a training-basedmethodwithneuralnetworkandunsupervised 108