Seite - 88 - in Document Image Processing

J. Imaging 2018,4, 41 gradients) [5],SIFT(scale-invariant feature transform)[6,7],LBP(localbinarypattern) [8]andSURF (speededuprobust features) [9]. Theseareprominent featureextractionmethods,whichhavebeen experimentedformanyproblemslike imagerecognition,character recognition, facedetection,etc. and thecorrespondingmodelsarecalledshallowlearningmodels,whicharestillpopular for thepattern recognition. Featureextraction[10] isone typeofdimensionalityreductiontechniquethat represents the importantpartsofa large image intoa featurevector. These featuresarehandcraftedandexplicitly designedbytheresearchcommunity. Therobustnessandperformanceof these featuresdependon theskill andtheknowledgeofeachresearcher. Thereare thecaseswheresomevital featuresmaybe unseenbytheresearcherswhileextractingthe features fromthe imageandthismayresult inahigh classificationerror. Deep learning inverts theprocessofhandcraftinganddesigningfeatures foraparticularproblem into an automatic process to compute the best features for that problem. A deep convolutional neuralnetworkhasmultipleconvolutional layers toextract the featuresautomatically. Thefeatures are extracted only once inmost of the shallow learningmodels, but in the case of deep learning models,multipleconvolutional layershavebeenadoptedtoextractdiscriminatingfeaturesmultiple times. This isoneof the reasons thatdeep learningmodelsaregenerally successful. TheLeNet [4] isanexampleofdeepconvolutionalneuralnetworkforcharacter recognition.Recently,manyother examplesofdeep learningmodels canbe listedsuchasAlexNet [3],ZFNet [11],VGGNet [12] and spatial transformernetworks[13]. Thesemodelshavebeensuccessfullyappliedforimageclassification andcharacterrecognition.Owingtotheirgreatsuccess,manyleadingcompanieshavealso introduced deepmodels. GoogleCorporationhasmadeaGoogLeNethaving 22 layers of convolutional and poolinglayersalternatively.Apartfromthismodel,Googlehasalsodevelopedanopensourcesoftware librarynamedTensorflowtoconductdeep learningresearch.Microsoftalso introduceditsowndeep convolutional neural network architecture namedResNet in 2015. ResNet has 152-layer network architectureswhichmade a new record in detection, localization, and classification. Thismodel introducedanewideaof residual learning thatmakes theoptimizationand theback-propagation processeasier thanthebasicDCNNmodel. Character recognition isafieldof imageprocessingwhere the image is recognizedandconverted into amachine-readable format. As discussed above, the deep learning approach and especially deep convolutional neural networks have been used for image detection and recognition. It has also been successfully applied onRoman (MNIST) [4], Chinese [14], Bangla [15] andArabic [16] languages. In thiswork,adeepconvolutionalneuralnetwork isappliedforhandwrittenDevanagari characters recognition. Themaincontributionsofourworkcanbesummarized in the followingpoints: 1. Thiswork is thefirst toapply thedeep learningapproachonthedatabasecreatedbyISI,Kolkata. Themaincontribution isarigorousevaluationofvariousDCNNmodels. 2. Deep learning is a rapidly developing field, which is bringing new techniques that can significantlyameliorate theperformanceofDCNNs. Since these techniqueshavebeenpublished in the last fewyears, there isevenavalidationprocess forestablishingtheir cross-domainutility. Weexploredtheroleofadaptivegradientmethods indeepconvolutionalneuralnetworkmodels, andweshowedthevariation inrecognitionaccuracy. 3. TheproposedhandwrittenDevanagaricharacterrecognitionsystemachievesahighclassification accuracy, surpassingexistingapproaches in literaturemainlyregardingrecognitionaccuracy. 4. A layer-wise technique of DCNN technique is proposed to achieve the highest recognition accuracyandalsogeta fasterconvergencerate. Theremainderofthispaperisorganizedasfollows. Section2discussespreviousworkinhandwritten Devanagari character recognition, Section 3 presents the introduction of deep convolutional neural networkandadaptivegradientmethods,Section4outlines theexperimentsanddiscussionsand, finally, Section5concludesthepaper. 88