Seite - 57 - in Document Image Processing

J. Imaging 2018,4, 57 In the literature,different texture featuredescriptorshavebeenusedtoseparate textandnon-text regions inprinteddocuments.Here,wehaveconsideredtwoof therecentonesandcomparedtheir individualperformancesonourdataset,with theperformanceof theRULBPoperator. Oneof the methodsusesGLCMasfeaturedescriptor [4]while theotherusesHistogramofOrientedGradients (HOG)[29]. Table4gives theaccuracyofclassificationforeachof the three featuredescriptorsusing allfiveclassifiers. It canbeseenthat theRULBPoperatoroutperformstheother featuredescriptors in mostcases. Table4.Performancecomparison in termsof recognitionaccuracy(in%)ofGLCM,HOGandRULBP (th=105)onthepresentdataset forfivedifferentclassifiers. Method NB MLP SMO KNN RF RULBP 50.38 90.78 88.62 90.20 91.96 GLCM 77.92 90.22 87.21 87.70 90.90 HOG 36.22 80.46 72.61 88.89 91.42 6.Conclusions In thepresentwork,ourobjective is tovalidate theutilityofLBPbasedfeaturedescriptors for the classificationof textandnon-textcomponentspresent inhandwrittendocuments, inacomprehensive way. We have experimentally shown that RLBP performs better than simple LBP, ILBP, RILBP, ULBPandRIULBP.However,amajor issueinusingRLBPistheselectionofasuitablethreshold,which mightbedomainspecific. In thecurrent researchattempt,wehaveselectedtheoptimalvalueof the thresholdonthebasisofa fewobservations,which isalsovalidatedthroughanexperiment.Wehave provideda justificationfor this selectionaswell,whichwebelievewould leadtodeeper insight into theselectionof the thresholdusedforLBP,especially in thecaseofhandwrittendocuments. Excluding that,wehaveproposedaminormodification toRLBPby incorporating the concept of a ‘uniform pattern’ todevelopRULBP,andithasbeenshownexperimentally thatRULBPperformsbetter than RLBP. In the future,wewould lookfor theother texturebasedfeaturesalongwithsomeothervariants ofLBP to see theirutility in the current context. In the future,weplan to enlarge thedatabaseby incorporatingvarious typesofdocument images,which, in turn,wouldmotivatemoreresearchers todosometangiblework. It isworthmentioninghere that, inorder toanalyze the textswritten in differentscripts, ascript recognitionmodule is required[30], sinceanOCRengine is script specific. Thus,our futureplan is to incorporate thesameinourmodel tomake itmoreuseful inamulti-script environment.Anotherarea thatwewill look into is thegeneralizationof the thresholdvalue th, so that wemayformulateasolidsetofprocedures thatcanbeuseful foranydocument, insteadofusingan empiricalmethodtodetect thesame. Acknowledgments:Thiswork ispartiallysupportedbytheCenter forMicroprocessorApplicationforTraining EducationandResearch(CMATER)research laboratoryof theComputerScienceandEngineeringDepartment, JadavpurUniversity, India, andPURSE-II andUPE-II JadavpurUniversity projects. RamSarkar is partially fundedbyaDSTgrant (EMR/2016/007213). AuthorContributions:Thefirst threeauthors—SouravGhosh,DibyadwatiLahiriandShowmikBhowmik—have contributedequallytothepaper. ErginaKavallieratouandRamSarkarprovidedessentialguidanceandcorrections atvariousstagesof thework. Conflictsof Interest:Theauthorshavenoconflictof interest. Abbreviations Thefollowingabbreviationsareusedin thismanuscript: LBP LocalBinaryPattern GLCM Gray-LevelCo-OccurrenceMatrix CC ConnectedComponents BB BoundingBox 57