Seite - 154 - in Document Image Processing

J. Imaging 2018,4, 39 2.2.HistogramofOrientedGradients (HOG) HOGdescriptor [31]countsoccurrencesofgradientorientation in localizedportionsofan image whichwasfirstproposed forpedestriandetection in steady images. Theessential thoughtbehind theHOGdescriptor is that localobjectappearanceandshapewithinan imagecanbedescribedby thedistributionof intensitygradientsoredgedirections. Atfirst, thevaluesof themagnitudeand directionofall thepixels foreachof thewordimagesarecalculated.Next, eachpixel ispigeonholed in certain category according to its directionwhich is knownas orientationbins. Then, theword image isdividedinton (heren=10)connectedregions, calledcellsandforeachcell, ahistogramof gradientdirectionsoredgeorientations iscomputedfor thepixelswithin thecell. Thecombinationof thesehistogramsthenrepresents thedescriptor. Since thenumberoforientationbins is takenas8 for thepresentwork,an80-D(i.e., 10×8) featurevectorhasbeenextractedusingHOGdescriptor [30]. ThemagnitudeanddirectionofeachpixelofasamplehandwrittenTeluguwordimagearealsoshown inFigure4. (a) (b) (c) Figure 4. Illustration of: (a) handwritten Telugu word image, (b) its magnitude part and (c) its directionpart. 2.3.ModifiedLog-GaborFilterTransform(MLGTransform) Modifiedlog-Gaborfilter transform-basedfeatures,proposedinReference [20],hadperformed well in thescriptclassificationtaskandthereforearealsochosenasoneof the featuredescriptorsof ourproposedmethodology inorder to identify thescriptof theword images. Inorder topreserve the spatial information,aWindowedFourierTransform(WFT) isconsideredinthepresentwork.WFT involvesmultiplicationof the imageby thewindowfunctionand the resultantoutput is followed byapplying theFourier transform.WFTisbasicallya convolutionof the imagewith the low-pass filter. Since for textureanalysis,bothspatialandfrequencyinformationarepreferred, thepresentwork tries toachieveagoodtrade-offbetweenthese two.Gabor transformsuseaGaussianfunctionas the optimally concentrated function in the spatial aswell as in the frequencydomain [32]. Due to the convolutiontheorem, thefilter interpretationof theGabor transformallowstheefficientcomputation oftheGaborcoefficientsbymultiplicationoftheFourier transformedimagewiththeFourier transform of theGaborfilter. The inverseFourier transformis thenappliedon the resultantvector toget the outputfiltered images. The images,after lowpassfiltering,arepassedas input toa function thatcomputesGaborenergy feature fromthem.The input image is thenpassedtoafunctiontoyieldaGaborarraywhich is the arrayequivalentof the imageafterGaborfiltering. Thefunctiondisplays the imageequivalentof the magnitudeandtherealpartof theGaborarraypixels. For thepresentwork,bothenergyandentropyfeatures [33]basedonModifiedlog-Gaborfilter transformhavebeenextractedfor5scales(1,2,3,4and5)and6orientations(0◦, 30◦, 60◦, 90◦, 120◦and 150◦) tocapturecomplementary informationfoundindifferent scriptword images.Here, eachfilter is convolvedwith the input imagetoobtain60different representations (responsematrices) foragiven input image. Figure5showsoutput images formedafter theapplicationofModifiedlog-Gaborfilter transformforasamplehandwrittenBanglawordimage. 154