Page - 152 - in Document Image Processing

J. Imaging 2018,4, 39 Themain contributionof thepresentwork is the comprehensive evaluationof themajor classifier combinationapproacheswhichareeitherrulebasedorapplyasecondaryclassifier for information fusion. Themotivation is to improvetheclassificationaccuracyat theword-levelhandwrittenscript recognitionbycombiningtheresultsof thebestperformingclassifieronthreepreviouslyusedfeature sets. It isamulti-classclassificationproblemandin thepresentcase,12officiallyused Indic-scriptsare consideredwhichare:Devanagari,Bangla,Odia,Gujarati,Gurumukhi,Tamil,Telugu,Kannada,Malayalam, Manipuri,UrduandRoman. Threedifferent setsof featurevectorsbasedonbothshapeandtexture analysishavebeenestimatedfromeachof thehandwrittenwordimages. Identificationof thescripts inwhich theword images arewritten, is donewith these featurevaluesby feeding the same into differentMLPclassifiers. Soft-decisionsprovidedbythe individualclassifiersare thencombinedusing anarrayof classifier combination techniques. This kindofwork is implemented for thefirst time assumingthenumberof Indicscriptsundertakenandtherangeofcombinationtechniquesapplied. Thesystemdevelopedfor thescript recognitiontaskhere, isapartof thegeneral frameworkwhere different featuresetsandclassifieroutputscanbemodelledintoasinglesystemwithoutmuchincrease in thecomputation involved. Blockdiagramof thepresentwork isshowninFigure1. Figure1.Schematicdiagramof theproposedmethodology. 2. FeatureExtraction In thispaper, threepopular featureextractionmethodologieshavebeenusedfor thecombination namely,EllipticalFeatures [21],HistogramofOrientedGradients (HOG)[30]andModifiedlog-Gabor filter transform[20]. Thefirst featureset isapplied tocapture theoverall structurepresent in thescript wordimageswhereas therest twofeaturesetsdealwith the textureof thesame. These featureshave alreadyprovidedsatisfactoryresults to thischallengingtaskofhandwrittenscript identification. 2.1. EllipticalFeatures Thewordimagesaregenerally foundtobeelongated innaturewhichcanbettercoveredbyan ellipse. That iswhy;elliptical featuresareextractedfromthecontourandthe local regionsofaword image so that it is easier to isolate aparticular script. Twomore important notationsused in this subsectionare: (a)Pixel ratio (Pr)and(b)Pixelcount (Pc). Pr isdefinedas theratioof thenumberof contourpixels (object) to thenumberofbackgroundpixelsandthepixel countwhereasPc isdefined as thenumberofcontourpixels. Thefeaturesaredescribed indetail: 2.1.1.MaximumInscribedEllipse Theheightandwidthof theboundingboxarecalculatedforeachwordimage.Arepresentative ellipse is theninscribed(consideringtheorientationof theellipse) inside thisboundingboxhaving 152