Seite - 89 - in Document Image Processing

J. Imaging 2018,4, 41 2. PreviousWork Devanagarihandwrittencharacterrecognitionhasbeeninvestigatedbydifferentfeatureextraction methodsanddifferentclassifiers. Researchershaveusedstructural, statisticalandtopological features. Neuralnetworks,KNN(K-nearestneighbors), andSVM(Supportvectormachine)areprimarilyused forclassification.However, thefirst researchworkwaspublishedbyI.K.SethiandB.Chatterjee [17] in 1976. The authors recognized thehandwrittenDevanagari numerals by a structured approach whichfoundtheexistenceandthepositionsofhorizontalandvertical linesegments,D-curve,C-curve, left slant andright slant. Adirectional chain codebased feature extraction techniquewasusedby N.Sharma[18].Aboundingboxofacharactersamplewasdividedintoblocksandcomputed64-D directionchaincode features fromeachdividedblock, and thenaquadratic classifierwasapplied for therecognitionof11,270samples. Theauthors reportedanaccuracyof80.36%forhandwritten Devanagari characters. Deshpande et al. [19] used the same chain code featureswith a regular expressiontogenerateanencodedstringfromcharactersandimprovedtherecognitionaccuracyby 1.74%.Atwo-stageclassificationapproachforhandwrittencharacterswasreportedbyS.Arora [20] where sheused structural properties of characters like shirorekha and spine in thefirst stage and in another stage used intersection features. These features further fed into a neural network for theclassification. Shealsodefinedamethodforfindingtheshirorekhaproperly. Thisapproachhas been testedon50,000 samples andobtained89.12%accuracy. In [21], S.Arora combineddifferent features suchaschaincodes, four sideviews, andshadowbased features. These featureswere fed intoamultilayerperceptronneuralnetworktorecognize1500handwrittenDevanagari charactersand obtain89.58%accuracy. Afuzzymodel-basedrecognitionapproachhasreportedbyM.Hanmandlu[22]. The featuresare extractedbytheboxapproachwhichdividedthecharacter into24cells (6×4grid), andanormalized vector distance for each boxwas computed except the empty cells. A reuse policy is also used to enhance the speed of the learning of 4750 samples and obtained 90.65% accuracy. Thework presented in [23] computed shadow features, chain code features and classified the 7154 samples usingtwomultilayerperceptronsandaminimumeditdistancemethodforhandwrittenDevanagari characters. Theyreported90.74%accuracy.Kumar [24]has testedfivedifferent featuresnamedKirsch directionaledges, chaincode,directionaldistancedistribution,gradient,anddistance transformon the25,000handwrittenDevanagaricharactersandreported94.1%accuracy.Duringtheexperiment, he found the gradient feature outperformed the remaining four featureswith the SVMclassifier, andtheKirsch directional edges featurewas theweakest performer. A newkind of featurewas also created that computed totaldistance in fourdirectionsafter computing thegradientmapand neighborhoodpixels’weight from thebinary imageof the sample. In thepaper [25], Pal applied themeanfilter four timesbefore extracting thedirectiongradient features thathavebeen reduced using theGaussianfilter. Theyusedmodifiedquadratic classifieron36,172 samples andreported 94.24%accuracyusingcross-validationpolicy. Pal [26]has furtherextendedhisworkwithSVMand MILclassifieronthesamedatabaseandobtained95.13%and95.19%recognitionaccuracyrespectively. Despite the higher recognition rate achieved by existing methods, there is still room for improvementof thehandwrittenDevanagari character recognition. 3.DeepConvolutionalNeuralNetworks (DCNN) Thedeepconvolutionalneuralnetworkcanbebroadlysegregated into twomajorpartsasshown inFigure1, thefirstpart contains thesequenceofalternativeconvolutionalwithmax-pooling layers, andanotherpartcontains thesequenceof fullyconnected layers.Anobject canberecognizedbyits featureswhicharedirectlydependentonthedistributionsofcolorintensityintheimage. TheGaussian, Gabor, etc. filtersareusedtorecordthesecolor intensitydistributions. Thevaluesofakernel for these filtersarepredefined,andtheyrecordonlythespecificdistributionofcolor intensity. Thekernelvalues arenotgoing to changeasper the responseof theappliedmodel. However, inDCNN, thevalues of thekernelarebeingupdatedaccording to theresponseof themodel. Thathelps tofindthebest 89