Seite - 153 - in Document Image Processing

J. Imaging 2018,4, 39 majorandminoraxesequaltothewidthandtheheightoftheboundingboxandthecentreofanellipse isalso thecentreof thecorrespondingboundingbox. Thisellipsedivides thewordimage intoeight regionsRi, I=1,2 . . . , 8. Theboundingboxalongwith the inscribedellipse forahandwrittenBangla word imageare shown inFigure2b. Taking thevaluesofPr fromtheseeight regions, as shown in Figure2a,eight features (F1–F8) foreachhandwrittenwordimageareestimated.Now,another typeof feature,PcalongN(N=8for thepresentwork) linesparallel tomajor/minoraxisof therepresentative ellipsearecomputed. Themeanandstandarddeviationof thevaluesofPcalongmajor/minoraxisare takenas fouradditional features (F9–F12). Figure 2. Illustration of fitting (a) an imaginary ellipse inside theminimumboundary boxwhich dividesaBanglahandwrittenwordimage in8regionsasshownin(b). 2.1.2. Sectional InscribedEllipse Eachof thewordimagessurroundedbytheminimumboundingbox isagaindividedinto four equalrectanglesandarepresentativeellipseisfit intoeachoftheserectanglesusingthesameprocedure asdescribed in theprevioussubsection. Asa result, everyellipseproduceseight regions inside its rectangular area namely,Rij where 1≤ i ≤ 4 and1≤ j≤ 8whichmakes 8× 4= 32 regions in total. Atotalof32 featurevalues (F13–F44)using thePrvalues is computedfromthe32ellipses in similar fashion. 2.1.3.ConcentricEllipses These featurevaluesarecomputedbytakingtheentire topologyof thewordimage.Aprimary ellipse ismadecircumscribingthewordimagewithcentre takentobethemidpointof itsminimum boundingbox. Thevaluesof themajor andminor axesof the ellipse are taken into consideration. Afterfittingtheprimaryellipse, threeconcentricellipsesaredrawninside theprimaryellipsehaving thesamecentrepointas theprimaryellipseandmajorandminoraxesequal to1/4th,2/4thand3/4th ofmajor andminoraxesof theprimaryellipse respectively. These four ellipsesdivideeachof the wordimages into fourregions-Re1,Re2,Re3 andRe4. Thepartitioningof the fourregionsonasample handwrittenDevanagariwordimage is showninFigure3. Fromthefourregions, four featuresvalues (F45–F48)consideringthePr’sandfourfeaturevalues(F49–F52)consideringthePc’sof theregionsRe1, Re2,Re3 andRe4areestimated. Theremainingsixfeatures(i.e.,F53–F58)aretakenasthecorresponding differencesof thePr’sandPc’sbetweentheregionsRe1 andRe2,Re2 andRe3,Re3 andRe4 respectively. Theelliptical features (F1–F58)aresuitablynormalizedbytheheightandwidthof thecorresponding wordimage. Figure3.Figureshowingtheellipticalpartitionof fourregionsonasamplehandwrittenDevanagari wordimage. 153