Deep Learning in Text Recognition and Text Detection : A Review

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

Deep Learning in Text Recognition and Text Detection : A Review

Brindha Muthusamy1 , Kousalya K2

1Dept. of Computer Science and Engineering, Kongu Engineering College, Erode 3Professor, Dept. of Computer Science and Engineering, Kongu Engineering College, Erode ***

Abstract Detecting text in natural situations is a difficult task that is more difficult than extracting the text from those natural images in which the background and foreground are clearly separated and every character is isolated from the images. Text in the landscapes which is nature may occur in a range of states like a text in dark with the background light and vice versa, with a broad diversity of fonts, even for letters of the same word, sections of words can be overlapped by environment objects, making detection of these parts impossible. Deep learning which is a subset of Machine learning employs a neural network, a technique that replicates how the brain analyses data. An Optical Character Recognition engine has two parts: i)Text recognition and ii)Text detection. The process of locating the sections of text in a document is known as text detection. Since the different documents (invoices, newspapers, etc.) have varied structures, this work has historically proved difficult. A text recognition system, on the other hand, takes a portion of a document containing text (a word or a line of text) and outputs the associated text. Both text detection and text recognition have shown considerable promise with deep learning algorithms.

Keywords

: Deeplearning,ConvolutionalNeuralNetwork,textdetection,textclassification,OpticalCharacterRecognition, optimization

I. INTRODUCTION

Deeplearningalgorithmslearnabouttheimagebypassingthrougheachneuralnetworklayer.Forapplicationslike machinetranslationandimagesearches,OCRmechanismisbeingusedtorecognisetextwithinphotographs.Themethodfor recognisingthetext/charactersfromaphoto,sendsthefeaturesextractedtextfromanimage/photototheclassifierwhichis trainedtodistinguishindividualcharactersthatissimilartoobjectrecognition.Recognitionoftext,ontheotherhand,looksat thetextasagroupofmeaningfulcharactersratherthansinglecharacters.Atextstringcanberecognizedbyclusteringthe similar characters, which means that every character in the text must separate for recognition [21] Alternatively, in the instanceofclassification,trainthenetworkonalabelleddatasetsinordertocategorizethesamplesinthedatasets.

ConvolutionalNeuralNetwork(CNN)wasappliedtoextractthefeaturesforcharacterrecognitionperformance.Abrief reviewontextdetectionandtextrecognitionareshowninTable2.TheCNNmodelistrainedinthreesteps:(i)TraintheCNN modelfromthescratch,(ii)Usingatransferlearningmethodforexploitingthefeaturesfromapretrainedmodelonlarger datasetsand,(iii)Transferlearningandfine tuningtheweightsofanCNNarchitecture.CNNarchitectureconsistsmostlyof fourlayers.(i)Convolutionallayer[Conv];(ii)Poolinglayer;(iii)Fullyconnectedlayer;and(iv)RectifiedLinearunits [4] ArtificialNeuralNetwork(ANN)isbeingtrainedtoextractinformationthroughthedeep learning basedtechnologies fromanimage.VGG,ResNet,MobileNet,GoogleNet,Xception,andDenseNetareexamplesofCNNarchitecturewhereseveral convolutionlayersareutilizedtoextractthefeaturesfromtheimages [21].Handwrittentextsaredifficultwhilereadingtext fromphotographsorrecognisingthathasattainedalotofattention.Mostsystemshavetwoimportantcomponents,(i)Text detection;and(ii)Textrecognition.Textdetectionisatechniquewherethetextinstancesfromtheimagescanbepredicted andlocalized.Thetextrecognitionisdonebyautoencoderwhichistheprocessofdecodingthetextintothemachine readable format.

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Figure1CNNarchitecturefortextclassification

II. TEXT DETECTION AND RECOGNITION USING DEEP LEARNING

2.1 Text detection

ThelocalfeaturescanbeextractedfromCNNthroughthetrainingofcharacterpicturesandsub regionsbasedonthe characteristicsofanindividual’shandwriting.TotrainCNN,randomlyselectedinstancesofanimageinthetrainingsetsare used,andthelocalfeaturesextractedfromtheimagefromtheseinstancesareaggregatedinordertogeneratetheglobal features.Theprocessofrandomlysamplinginstancesisrepeatedforeverytrainingepoch.Thisresultsintheincreaseof trainingthepatternsfortrainingtheCNNfortextindependentwriteridentification [1].

Deep learning can constantly learn by examining data and finding patterns and classifying images. For picture categorization,languagetransactionandcharacterrecognition,deeplearningisapplied.DeepNeuralNetworksarenetworks withmultiplelayersthatcanperformcomplexoperationsonimages,sound,andtext,suchasrepresentationandabstraction.It canbeusedforanytypeofrecognitionproblem.Thebasicgoalisonlyforcomputerstolearnwithouthumaninterventionand modifytheiractivitiesaccordingly.

2.2 Text recognition

Therecognitioncanbedonebymanytechniques.ItinvolvesConvolutionalNeuralNetwork(CNN),SemiIncremental Method,IncrementalMethod,LineandWordSegmentationmethodetc.Oneofthe mosteffectiveandprominentwaysof handwritingrecognitionisConvolutionalNeuralNetwork(CNN).ThemostprevalentapplicationofCNNisinimageanalysis. ArtificialneuronsareusedinCNN [11].CNNcanrecogniseimagesandvideos,classifyimages,analysemedicalimages,perform computervision,andprocessnaturallanguage.

CNNorConvNetsarchitecturehasmadesignificantcontributionstotheanalysisofimages.CNNisdefinedas,1)A convolutiontoolthatseparatesandidentifiesthedistinctcharacteristicsoftheimageforanalysisinaprocesscalledFeature Extraction,whichispartoftheCNNarchitecture.2)Afullyconnectedlayermakesuseoftheoutputoftheconvolutionprocess inordertoforecasttheimage'sclassusingtheinformationacquiredinearlierstages.

CNNextractsthefeaturesfromhandwrittentextimagesofnumerouscharactersbytheend to endmethodbasedon deeplearningandcombinestheseextractedfeaturestorecognizethewriter’sspecificdataduringloweringofcharacter’sclass specific features.Tocreaten tupleimages,asinglewriter’shandwrittensquaredimagesaresampledrandomly.Second,every imagefromn tupleissentintoalocalfeatureextractor(CNN).CNNcanextracttext independentwriterspecificpropertiesby employingnewtechniquestostructurethetrainingsamplesasntuplepictures.Finally,aglobalfeatureaggregatoraggregates theretrievedlocalfeaturesinvariousways,suchasusingtheaverageormaximum.Finally,thecombinedcharacteristicsare sentintoasoftmaxclassifier(fullyconnectedwithanNfcunitwhichisequaltothenumberofwriters)forpredicting [1]

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal |

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

III. PRELIMINARY STEPS FOR TEXT DETECTION AND RECOGNITION

3.1 Datasets

ThedatasetsarecollectedfromvariousImageNetdatasetlike,MNIST,NIST,IAMdataset,IFN/ENIT,andsoon.The different dataset on different languages like, English, Arabic, Bangla handwritten character datasets are taken into consideration.Thedetailedviewofthesedatasetsareshownintable2.

3.2 Image Pre-processing

Pre processingisatechnologywhichtransformstherawdatatousefulandefficientdata/information.Thisprocess consistsofdifferentoperationswhichareusedtoperformoninputimagesandinthisprocess,theimagesarereshaped.The rearrangementoftheformofthedataisdonewithoutchangingthecontentsofthedata.Differentkindsofarrangementsare doneinthisprocessaccordingtotheparameterswhichareneededtocarryontillfurtherprocess [1].Imagepre processingis nothingbuttoremoveirrelevantdataanddeletetheduplicatedatafromthedatabases.Thetextintheimagecanbeindifferent styles,fontsandsoon.Indatapre processing,thedataareprocessedindifferentways,theyare:



Data cleaning :Removenoiseanddatainconsistencyfromrawdata.Thatis,the datasetwhichdoesnotbelongtothe datasetcomes,thisdata cleaningprocessremovesthesedatasets. It isdone byfiltering the datasetandhandlingthe missingdata.Itensuresthequalityofdata.Inotherwords,itonlyremovestheunwanteddatabutmaintainstheoriginality ofthedata.



Data integration :Dataintegrationcollectsdatafromvariousmultipleresourcesandcombinesittoformcoherent dataandalsosupportstheconsolidatedperspectiveoftheinformation.Thatis,itmergesdatafromvarioussourcesintoa coherent datastore (data warehouse). These data are stored and maintained for future use. It may be in the form of documentswheretherelevantdataarestoredinthedifferentdocuments.



Data reduction : Data reduction is the process of reducing the data size by instances, aggregation, eliminating irrelevantdata/featuresorbyclustering.Datareductioncanincreasethestoragecapacityandthecostisreduced.Itdoes notloseanydata,insteaditmaintainstheoriginalityofthedata.Datareductionprocessisdonewiththehelpofdata compression.Datacompressionisatechniquetocompressthedatainareducedformofdatawhichcanreducethestorage space.



Data transformation :Datatransformationisusedtoconvertoneformofformattoanotherformat.Itisalsoknown asdatamungingordatawrangling.OtherwisecalledNormalization.Thedatapointsinthescatterplotshouldbelinear.If thepointsareintheformofacurve,itisdifficulttocalculatetheaccuracywhichaffectstheperformanceofthemodel.This curvecanbeconvertedintoalinearlinebyscalingthemodelintherangeof0and1.Activationfunctionsareusedwhen themodelshows non linearityintheirrespectivemodel.

3.3 Normalization

Normalizationistheprocessoforganizingdatainthedatabase.Thisprocessinvolvestablecreationandestablishing relationshipsbetweenthesetables.Itprotectsthedataaswellasmakesthedatabasemoreflexiblebyeliminatingredundancy andinconsistentdependency.Redundantdatatakesmorespaceandiscomplicatedtomaintain.Theinconsistentdependent datacanmakeitdifficulttoaccessthedatasincethepathtofindthedataiseithermissingorbroken.Hence,Normalizationis moreimportantsinceitcanreducetheseirrelevantdataanddatainconsistency anditcanhandlethemissingdataaswellto makethedatabasemoreflexible.Normalizationincludesthreestagesofnormalizationstepswhereeachstagegeneratesthe table.Eachtablestorestherelevantdatawhichdoesnotincludeduplicatedataormissanydata.Thethreestepsinvolvedin thenormalizationprocessareasfollows,

3.3.1 First Normal form

TheFirstNormalForm(1NF)setsthefundamentalrulesfordatabasenormalizationwhichrelatestothesingletablein therelationaldatabasemodel.ThestepsinvolvedintheFirstNormalFormare:

Everycolumninthetableareunique

Separatetablesarecreatedforeveryrelevantsetofdata

Eachtablemustbeidentifiedwithauniquecolumnortheconcatenatedcolumnsarecalledwiththeprimarykey

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

Neitherroworcolumnofthetableareduplicated 

NoroworcolumnthatintersectsinthetablecontainaNULLvalue 

Noroworcolumnthatintersectsinthetablecanhavemulti valuedfields

3.3.2 Second Normal Form

TheSecondNormalForm(2NF)followstheFirstNormalForm.ThebasicrequirementsofSecondNormalFormfor organizingthedatainclude: 

Noredundancyofdata.Allthedataisstoredinonlyoneplace. 

DatadependenciesinSecondNormalFormarelogical.Thatis,alltherelateddataitemsarestoredtogetherwhichis usefulforeasyaccess.

3.3.3 Third Normal Form

TheThirdNormalForm(3NF)isthecombinationofbothFirstNormalFormandSecondNormalForm(1NF+2NF).The mainbenefitsofthisThirdnormalformare: 

Reducesduplicationofthedataandachievesdataintegrityinadatabase. 

Usefultodesignanormalrelationaldatabase 

3NFareindependentofanomaliesofdeletion,updation,andinsertion 

Itensureslosslessnessandpreventionofthefunctionaldependencies

3.4 Feature extraction

TextextractionorFeatureextractionfromanimageisamethodofextractingtextfromaphotographusingmachine learningtechniques.Textextractionisalsoknownastextlocalization.Itisusedfortextdetectionandlocalizationwhichhelps fortextrecognition.Textlocalizationistheprocesswhichisusedtodevelopacomputersystem(AI)toautomaticallyrecognize andreadthetextfromtheimages.Indeeplearningmodels,thefeatureextractionprocessisdoneautomaticallysincethe modelsarepre trained.Itisdonebytextclassification.

3.4.1 Text Classification

Textclassificationisalsoknownastexttaggingortextcategorization.Textsarealwaysunstructuredinhandwritten wordsorcharacters.Hence,thesetextsarecategorizedintoanorganizedgroup.Thisprocessisdifficultsinceittakesmore timeandisalittleexpensive.ByusingNaturalLanguageProcessing(NLP)whichcanbeusedasatextclassifier,thetextscan be categorized and can be converted into structural textual data which is easy to understand, cost effective and is more scalable.NaturalLanguageProcessingcanautomaticallyanalyseandunderstandthetypeofcharacteranditwillassignasetof pre definedtags(Pre trainedimagedatasets)andcanbecategorizedbasedonitscharacteristics.

3.5 Text detection

Inthetestingstage,textdetectionisaccomplishedbysegmentingphotos.Aseriesoftext/sub imagesofindividualtext isdividedfromafullimage.Edgedetectionandthespacebetweenthedifferentcharactersareusedtosegmenttheimage. Followingsegmentation,thesub dividedportionsarelabelledandprocessedoneatatime.Thislabellingisusedtodetermine thetotalnumberofcharactersinanimage.Afterthat,eachsubpictureisscaled(70x50)andnormalizedinrelationtoitself. Thisaidsintheextractionofimagequalityattributes [9].Textdetectionisamethodinwhichthemodelisgivenanimageand thetextregionisdetectedbybuildingaboundingboxaroundit.Textrecognitioniscarriedoutbyfurtherprocessingthe discoveredtextualsectionsinordertorecognisethetext.

3.5.1 Training the dataset

Deeplearningmodelisbuiltwhilefeedingdatatoadeepneuralnetwork(DNN)to"train"inordertodoaspecificAI task (such as image classification or speech to text conversion). The hidden layer in the neural network is used for backpropagationwhichhelpstoimprovetheperformanceofthetrainingmodel.Hence,duringthetrainingprocess,known dataisfedintotheDNN,andthisDNNgeneratesapredictiononwhatthedatarepresents.

2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page12

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

ImageClassification(usedtoclassifythetypeofanobjectinanimage).

o Input:Aphotograph,forexample,isasingle objectimage.

o Output:Adesignationforaclass(oneormoreintegersthataremappedtoclasslabels).

3.5.2 Validating the dataset

Validationdataintroducesnewdataintothemodelthatithasn'tassessedbeforeduringtraining.Validationdata servesasthefirsttestagainstunknowndata,allowingdatascientiststoassesshowwellthemodelpredictsnewdata.Although notalldatascientistsusevalidationdata,itcanbeusefulinoptimizinghyperparameters,whichinfluencehowthemodel evaluatesdata.

ObjectLocalization(locatetheexistenceofobjectsinanimageanduseaboundingboxtorepresenttheirlocation).

o Input:Aphotograph,forinstance,isanimagefeaturingoneormoreobjects.

o Output:Asetofboundingboxes(ormore).

3.5.3 Testing the dataset

Testingdataoncethemodelhasbeendevelopedconfirmsthatitcanmakeaccuratepredictions.Thetestingdata shouldbeunlabelledifthetrainingandvalidationdataincludelabelstotrackthemodel'sperformancemetrics.Testdataisthe verificationofanunknowndatasettomakesurethatthemachinelearningalgorithmwastrainedproperly.

ObjectDetection(detectthepresenceofobjects/charactersinanimagewithaboundingbox,aswellasthetypesor classesofthoseobjects).

o Input:Aphotograph,forinstance,isanimagefeaturingoneormoreobjects.

o Output:Aclasslabelforeachboundingbox,aswellasoneormoreboundingboxes(specifiedbyapoint,width,and height).

3.6 Text recognition

Textrecognition,commonlyknownasOCR(OpticalCharacterRecognition)isatypeofcomputervisionproblemwhich involvesconvertingimages/photosofdigitalorhand writtencharacterintoamachine readabletextthatthecomputercan process,save,andeditasatextfileoraspartofdataentryandmanipulationsoftware.Itcanrecognizethetextfromany documents,handwrittencharacters, andsoonbypre trainingtheimagesusingdeeplearningconceptsanditsarchitectures.In otherwords,text recognitionistheprocesswheretheimagetextisconvertedinto a recognisableandreadable text.The recognitionofthedetectedtextisdonebyvariousdeepneuralnetworks.

Year CNN architecture No. of layers No. of parameter

2014 VGG 16 16layers 138million

2015 ResNet 34layers 23million

2014 GoogleNet 22layers 7million

2016 Xception 71layers ~26million

DenseNet 121layers 20million

Table1CNNarchitecture

Some of the advantages of Optical Character Recognition

Quickdataretrieval

2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

Easytoextractrelevantdata

Lowcostforusagelikecopying,editingandsoon

Highaccuracytodetectandrecognize

Increasethestoragecapacity

IV. DEEP NEURAL NETWORKS FOR TEXT DETECTION AND RECOGNITION

Deeplearningisadomainofstudythataimstocopythefunctioningofhumanbrainsinordertoprocessdataand makedecisions.DeeplearningisalsoknownasDeepNeuralNetwork(DNN)orDeepNeuralLearning.Thedeeplearningmodel forOpticalCharacterRecognitionusestwotypesofneuralnetworkarchitecture.Theyarei)ConvolutionalNeuralNetwork (CNN);ii)ArtificialNeuralNetwork(ANN).

4.1 Convolutional Neural Network (CNN)

ConvolutionalNeuralNetwork(CNN)isaneuralnetworkwhichhasthreelayers.Theyare:(i)Inputlayer,(ii)Hidden layer;and(iii)Outputlayer.

Optical Character Recognition (OCR) considers the optical image of a character as an input and provides the correspondingrecognisablecharacterasanoutput.ThetrainedCNNisusedtoextractthefeaturesoftheimage.CNNclassifiers alongwithotherclassifierscanbecombinedandgivethebestresultforclassification.Thatis,theaccuracyandefficiencyofthe classificationcanbeimproved.Thehiddenlayerusestheinputtoperformfeed forwardandbackpropagationinorderto improve the accuracy and reduce the error rate. Convolutional Neural Network and Error Correcting Output Code (CNN +ECOC)whereCNNforfeatureextractionandECOCforclassificationarecombinedtoobtainOpticalCharacterRecognition. Thismethodgiveshighaccuracyforhandwrittencharacters.ItistrainedandvalidatedwithNISTdataset [46].

4.2 Artificial Neural Network (ANN)

Artificial Neural Network can be used to recognise the image with a single character and further include many characters for classification. The classification process can be done in two phases: i) Feature preprocessing read each characterandconvertitintobinaryimageandscanbyallfoursides(left,tp,right,bottom) andii)a)Trainingtheneural networkandb)Testingtheneuralnetworkwiththedatasets.Here,thetrainingsetsareusedtolearnhowtoremovenoise fromthedata [50]

4.3 Recurrent Neural Network (RNN)

RecurrentNeuralNetworkisbasedonthesequentialformofdata.Recurrentneuralnetworkismostsuitableforthe textclassificationwhichcanrecognizethewholesentenceorsequentialsetofwords.Whenthewordsorcharactersarein sequence,thisrecurrentneuralnetworkcanpredictthenextwordofthesequence.Hencehandwrittencharacterscanbe predictedmoreaccuratelywhileusingrecurrentneuralnetworks.Toachievethepredictionaccuracy,RNNdoesnotrequirea datasetintheformoflabelleddata(neednotbesupervisedlearning).Recurrentneuralnetworksarecapableofworkingthe temporalinformation. ThedisadvantageofRNNisthatitwillraisethevanishinggradientproblem.Thetrainingofdatasetsis morecomplexandhencemoredifficult.Itisalsodifficulttoprocessthelongsequentialcharacters.

LEARNING RATE IN DEEP LEARNING ALGORITHM

The learning rate refers to the number of times the images are trained. In deep learning neural networks, the StochasticGradientDescentisused.Thelearningratereferstotheparametersorhyperparameterwhichcontrolshowmany timesthedatasetistrainedinordertoreduceerrorrateandimproveaccuracywhichimprovestheperformanceofthemodel. Whiletrainingthisdataset,backpropagationtechniqueisusedwheretheweightsareupdatedeachtimeinordertogetbetter performanceoftheneuralnetworkmodel.Thislearningrateaffectstheperformanceofthemodelwhenitreachesthelocal minima.Hence,itisimportanttoadjustthelearningratefromhightolowtoslowdownoncethemodelreachestheoptimal solutionduringtraining.

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Figure2ProcessofCNNmodel

V. CHALLENGES IN TEXT DETECTION AND RECOGNITION

1. Awiderangeofnaturalphotos Charactershavedifferentfont styles,sizes,andcolours,aswellasdistinctivefont alignmentinthenaturalimages.Thelanguagealsovariesdependingonthestateorcountryorevenotherregion.

2. Backgroundcomplexity Naturalphotographs'backgroundsmaynotbeasdistinct.Thebackgroundcomprises grasses,bricks,pebbles,andsignboards,makingtextidentificationmoredifficult.

3. Factorsinfluencinginference Blurring,noise,andlowresolutionofinputimagesarethe keyinferencevariables. Thephotosmayhaveblurredtextthatcannotberecognizedandreadthetextaccurately.

4. Poorhandwriting Handwrittencharactersmaynotbeclearduetovariousstyleswhichmaynotbereadable(since handwritingvariesperindividual).

5. Sizeofcharacters Aperson’shandwritingvariesforeveryindividualwhichiswhysomeofthecharactersarenot easilyrecognisable.

6. Languageidentification Thetextintrafficsignsandadvertisingpanelsmaybeindifferentlanguagesthatarenot known by foreigners. The artificially added text on an image like watermarks or subtitles are also difficult to recognizeespeciallywhentheyareinforeignlanguage.

7. Thetextinthevideosareverydifficulttoreadsincethetextismovingcontinuouslyataparticularspeed.Someof thesetextsmayormaynotbeclearwhichisverydifficulttorecognizeaccurately.

VI. DEEP LEARNING ALGORITHMS

6.1 Convolutional Neural Network (CNN)

Toperformconvolutionaloperations,CNNsprocessinputbypassingitthroughmanylayersandextractingfeatures. TheRectifiedLinearUnit(ReLU)thatoutlaststofixthefeaturemapmakesuptheConvolutionalLayer.Thesefeaturemapsare rectifiedintothenextfeedusingthepoolinglayer.Poolingisadown sampledmethodwhichreducesthedimensionofthe featuremap.Theresulting2 Darraysaremadeupofsingle,long,continuous,andlinearvectorsthathavebeenflattenedinthe map.Fully connectedlayeristhenextlayer,whichtakesaflattenedmatrixor2 Darrayobtainedfromthepoolinglayerasan inputandthenclassifiestheimage.

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page15

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

6.2 Long Short Term Memory Networks(LSTMs)

LSTMsareproven whichoutperformstheconventional recurrent neural networksanditsuffersfromthefading gradientproblem,whenmodellinglong termdependencies.TheCTClayer,whichcomesaftertheLSTMlayers[figure3],gives thefeaturesequenceswiththegroundtruthtranscriptionduringthetraininganddecodingtheLSTMlayer’soutputsduring evaluationtocreatethepredictedtranscription.Fromthestarttofinish,thesystemistaughtthroughfeedingthetextlines, picturesandgroundtruthtranscription(inUTF 8) [5].

Figure3CNN BiLSTMarchitecture

6.3 Recurrent Neural Networks (RNNs)

Recurrent Neural Networks is mainly used for feed forward techniques. RNN can work both parallelly and sequentially.Duringcomputation.AnumberofsequencelearningchallengeshavebeensuccessfullysolvedusingRecurrent Neural Networks (RNN) and RNN variations (Bi directional LSTMs and MDLSTM). According to several studies, LSTM outperformsHMMsonsuchtasks.

6.4 Generative Adversarial Networks (GANs)

GANsaretheformofneuralnetworkarchitecturewhichenablesthedeeplearningmodels tolearnandcapturethe trainingdatadistribution,whichprovidesthegenerationofnewdatainstancesbasedonthesedistributions.GANsareusedfor unsupervisedmachinelearningtotrainthetwomodelsparallely.Itgivesthetrainingdatawhichissimilartotheoriginaldata. AdiscriminatorandageneratorareoftenincludedinGANs [51].

6.5 Multi Layer Perceptron (MLPs)

The MLPs is a supervised learning convolutional artificial neural network that reduces error by continuously computingandupdatingalltheweightsinthenetwork.Inthefirstphase,whichisafeed forwardingphase,thetraineddatais delivered to the output layer, after that the output and the desired targets (errors) are back propagated to update the network’sweightinthenextphase.TheAdamoptimiserwasemployedtoimprovetheperformanceofMultiLayerPerceptron [22].

Adam optimizer : Adamoptimizerisanextensionofstochasticgradientdescentthatusesdeeplearningapplications forcomputervisionandNaturalLanguageProcessing(NLP).Adamoptimizerisusedtoremovevanishinggradientproblems. Thecomputationtimeisfasteranditrequiresonlyafewparametersfortuning.

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

6.6 Deep Belief Networks (DBNs)

ADeepBeliefNetwork(DBN)isatypeofdeepneuralnetwork(DNN)oragenerativegraphicalmodel.TheDBNis constructedwithmultiplelayersof“hiddenunits”whicharelatentvariableswiththeconnectionsacrossthelevelsandnotin betweentheunitsineachlayer.DBNcanbelearnedtoprobabilisticallyrecreatetheinputswithoutsupervisionwhentheseare trainedonthesetofexampledatasets.Next,theDBN’slayersactasafeaturefordetection.Additionally,aDBNcanbetrained under supervision to categorize the text or object after completing the learning stage. To develop a Novel Q ADBN for handwrittendigit/characterrecognition,ADAEandQ learningalgorithmsareintroducedinDeepBeliefNetwork [4].

6.7 Restricted Boltzmann Machines (RBMs)

TheAdaptiveDeepAuto encoder(ADAE)hasbeenmadeupofnumeroussuccessivelystackedRBMs,wheretheoutput ofoneRBMactsastheinputofthenext.EachRBNistreatedasanencoder.Theencoderusesuniquecodetorecognisethetext .ADAE’shierarchicalfeatureextractiontechniqueiscomparabletothatofahuman’sbrain [4].

6.8 Autoencoders

Autoencoders are the combination of Encoders and Decoders. It is mainly used to learn the compressed representationsof thedatasets. The Autoencodersshouldbe trained in ordertolearnthefixeddimensional latentspace representationofthegivenimage,makingthemidealforthefeatureextraction [6]. Theoutputoftheautoencodernetworkis thereconstructionoftheinputdatawhichismoreefficient.

Related Works

Referen ce Proposed Algorithm

Purpose Dataset Accuracy

1

Convolutional Neural Network extractfeaturefromrawimages

JEITA HP database Firemaker and IAM database

99.97% 91.81% 2 Codebookmodel, Clustering,Bayesianclassifier, Moore’salgorithm

Feature generation and Feature selection, Classifiers using feature vectors

IAMdataset AUT FHdataset

93.7% 96.9% 3 Convolutional Neural Network+XGradientBoost CNN featureextraction XGBoost Recognition and classification

HECR (CNN+XGBoost) 99.84% 4 Q ADBN ADAE extractfeaturesfromoriginalimages usingADAE MNISTdataset 99.18% 5 Convolutional Neural Network, Long Short Term Memory network

featureextraction IncreasethememoryofRNN IFN/ENITdataset 83% 6 Convolutional Neural Network autoencoder + SupportVectorMachine

Featureextractor Classifytheimages 4600 MODI characters 99.3% 7 Convolutional Neural Network, Algebraic fusion of multiple classifier

Feature extractor Multi level fusion classifier MNISTdataset 98% 8 ArtificialNeuralNetwork ClassifyandRecognizethetextfrom USTB Vid TEXT 85%

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

images dataset 9 ArtificialNeuralNetwork Pre processingSegmentation,Feature Extraction 95% 10 Convolutional Neural Network RecurrentNeuralNetwork MultidimensionalLong Short TermMemoryNetworks

UPTIdataset 98.12% 11 Convolutional Neural Network FeatureExtractor

FeatureExtractor Normalizationandstandardisation

NMISTdataset Testing Acc 98.85% Training Acc 98.60% 12 Convolutional Neural Network Extracttheimage NISTdataset 86% 13 LexiconConvolutionalNeural Network, RecurrentNeuralNetwork

Detectcommonwords Extractandclassifytheimages IAMdataset 99% 14 Convolutional Neural Network recognize handwritten digits, characters ‘hpl tamil iso char’ dataset Training accuracy 95.16% Testingaccuracy 97.7% 15 Deep Convolutional Neural Network toextractfeaturesfromrawdata

CMATERdbdataset Digits 99.13% Alphabets 98.31% Characters 98.18% 16 Convolutional Neural Network recognize handwritten digits, characters BanglaLekha Isolated CMATERdb ISIdataset Mixeddataset

95.71% 98% 96.81% 96.40% 17 Deep Convolutional Neural Network toextractfeaturesfromrawdata Ancient Kanada documents 92% 18 Support Vector Machine, LinearRegression, K Nearest Neighbour, RandomForest, MultiNomialBayesclassifier

Classificationofimages Classificationandregression Classifydiscretefeatures

extractlocalfeatures extractglobalfeatures Chinese Wikipedia datasets 78%

IMDB SPAMdataset 85.8% 98.5% 19 RecurrentNeuralNetwork extractinformation IAMdatabase 91.70% 20 Convolutional Neural Network, Bi LongShortTermMemory networks

International Research Journal of Engineering and Technology (IRJET)

e ISSN: 2395 0056

21 Deeplearningtechnology Image Preprocessing, Symbol detection, Textrecognition

P&IDsymboldataset 97.89% 22 LinearRegression, Long Short Term Memory networks, MultiLayerPerceptron, DecisionTree

Extractfeatures Classifyandclassifytheimages

Twitter and Non Twitterdatasets LR 99.80% LSVM 99.78% MLP 99.12% DT 99.74% 23 SupportVectorMachine Signrecognition PSLdataset 80 90% 24 KNearestNeighbour, RandomForest, NaiveBayes, SupportVectorMachine

Proteindataset KNN 98.6% RF 98.5% NB 96.42% SVM 97.38% 25 ArtificialNeuralNetwork RandomForest, KNearestNeighbor,

Textfiltering

Sensor based SL dataset ANN 99% RF 99% K NN 98.5% 26

Pattern recognition, Image classification Clustering

Extremelearningmachine (Single hidden layer feedforwardneuralnetwork)

Uniformrandominitialization,Xavier initialization,ReLU initialization, Orthogonalinitialization

ISI kolkata Odai numerical, ISI kolkata Banglanumerical NIT RKL BanglaNumerical

96.65%, 96.65% 96.89% 97.75% 27

CAPTCHA Convolutional Neural Network

Breakingprocessandframework CASIA HWDB dataset 99.84% 28

Extreme Deep Convolutional NeuralNetworks, DeepNeuralNetworks, LinearRegression

Regularization , Normalization and Binarization

SDH2019.2, MNISTdataset 98.85% 29

SupervisedMachineLearning, SupportVectorMachine

Preprocessingandfeatureextraction fromimages

KVIS Thai OCR Dataset 74.32% 30

Convolutional Neural Network Character recognition and feature extraction Devanagari handwritten characterdataset

93.73% 31

Convolutional Neural Network Pre processing, Charactersegmentation, Recognition

Devanagari handwritten characterdataset

98.47% 32

SEG WI

Normalization,max pooling

Featureselection, featureextraction featuregeneration

IAM CVL IFN/ENIT Devanagari

IFN/ENITdataset 90.02%

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page19

97.27% 99.35% 98.24% 87.24% 33 Beta ellipticmodel, Codebook implementation model

International Research Journal of Engineering and Technology (IRJET)

e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

34 Sliding Convolutional Neural Network, Slice Convolutional Neural Network

35 FasterRConvolutionalNeural Network,RRecurrentNeural Network, Tree Shaped Deep Neural Network

36 NovelHybridnetwork, BoF framework, HMMs, K meansclustering

37 Encoder Decoder, Convolutional Neural Network, Bi LongShortTermMemory networks

38 CRecurrentNeuralNetwork, deep Bi Recurrent Neural Network, Connectionist TemporalClassification

39 Deep Convolutional Neural Network

Dataset Collection and Annotation Textrecognition

DatasetPreparationforDetection, CustomFeature, EvaluationPrediction, LigatureRecognition

Pre processing, FeatureExtraction,Classification, Overlapping

Decodingmechanism Featureextraction Sequenceanalysis

ShopSigndataset 85%

SDAidataset 95.20%

40 R Convolutional Neural Network ATR Deep Convolutional NeuralNetwork

Fine tune the Bi Long Short Term Memorynetworks, SequenceLabelling,Transcription, NetworkTraining

Pre processing, Characterclassification Trainingfromscratch, Featureextractor, Fine tunetheCNN

RelaxationConvolution, AlternateTraining

P KHATTdataset 99.95%

SVT IIIT5K IC03 IC13

IIIT5K SVT IC03 IC13

84.5% 85.4% 91.9% 91%

81.2% 82.7% 91.9% 89.6%

OIHACDB,AHCD 98.86% 99.98%

MNISTdataset ATR CNN Error rate : 0.254±0.014

Table2Methodsusedfortextdetectionandrecognition (ALiteratureReview)

VII. CONCLUSION AND FUTURE SCOPE

Deeplearninghasagreatlearningabilityandcanonlybenefitfromtheconclusion,scalartransformation,andbackground switches.Inrecentyears,deeplearning baseddetectiontechniqueshavebecomeapopularstudytopic.Thisstudypresentsa comprehensiveoverviewof deeplearning based detectionandrecognitionstrategiesthatmaytacklea varietyofsub problems,suchasocclusion,clustering,andlowerresolution,usingmultipleDeepNeuralNetworkmethodology(DNNs).The reviewinthisstudyisbasedontextdetectionfromhandwrittencharactersusing CNN architecture.Thestudy thengoes throughobjectdetection,facedetection,andothertypesofdetection.Thisreviewcanbeappliedtoadvanceddiscoveriesin neuralnetworksandothercomparablesystemsthatusedeeplearningforanydetectionandclassification,anditprovides usefulandcrucialguidelinesforfutureimprovement.Inthefuture,OpticalCharacterRecognitionwillplaya vitalroletofinda waytodigitisethewordsandnumbersinphysicallywrittentextandcharactersindifferentlanguages.

REFERENCES

1.Abdi,M.N.,Khemakhem,M.,2015.Amodel basedapproachtoofflinetext independentArabicwriteridentificationand verification.PatternRecognition48,1890 1903.

Factor value: 7.529 | ISO 9001:2008 Certified Journal

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

2. AlJarrah, M.N., Zyout, M.M., Duwairi, R., 2021. Arabic Handwritten Characters Recognition Using Convolutional Neural Network,in:202112thInternationalConferenceonInformationandCommunicationSystems(ICICS).IEEE.

3.Alom,M.Z.,Sidike,P.,Hasan,M.,Taha,T.M.,Asari,V.K.,2018.HandwrittenBanglaCharacterRecognitionUsingtheState of the Art Deep Convolutional Neural Networks. Computational Intelligence and Neuroscience 2018, 1 13. https://doi.org/10.1155/2018/6747098.

4.Alwajih,F.,Badr,E.,Abdou,S.,2022.WriteradaptationforE2EArabiconlinehandwritingrecognitionviaadversarialmulti tasklearning.EgyptianInformaticsJournal.https://doi.org/10.1016/j.eij.2022.02.007

5.Arafat,S.Y.,Iqbal,M.J.,2020.Urdu TextDetectionandRecognitioninNaturalSceneImagesUsingDeepLearning.IEEEAccess 8,96787 96803.https://doi.org/10.1109/access.2020.2994214

6. Boufenar,C.,Batouche,M.,2017.Investigationondeeplearningforoff linehandwrittenArabicCharacterRecognitionusing Theanoresearchplatform,in:2017IntelligentSystemsandComputerVision(ISCV).IEEE.

7. Breuel, T.M., n.d. Handwritten character recognition using neural networks, in: Handbook of Neural Computation. IOP PublishingLtd.

8.Chowanda,A.,Sutoyo,R.,Meiliana,Tanachutiwat,S.,2021.ExploringText basedEmotionsRecognitionMachineLearning Techniques on Social Media Conversation. Procedia Computer Science 179, 821 828. https://doi.org/10.1016/j.procs.2021.01.099.

9. Elkhayati, M., Elkettani, Y., 2022. UnCNN: A New Directed CNN Model for Isolated Arabic Handwritten Character Recognition.ArabianJournalforScience andEngineering.https://doi.org/10.1007/s13369 022 06652 5.

10. Ghiasi,G.,Safabakhsh,R.,2013.Offlinetext independentwriteridentificationusingcodebookandefficientcodeextraction methods.ImageandVisionComputing31,379 391.https://doi.org/10.1016/j.imavis.2013.03.002.

11. Guo,H.,Liu,Y.,Yang,D.,Zhao,J.,2021.OfflinehandwrittenTaiLecharacterrecognitionusingensembledeeplearning.The VisualComputer.https://doi.org/10.1007/s00371 021 02230 2.

12. Han,C.,n.d.NeuralNetworkBasedOff lineHandwrittenTextRecognitionSystem.FloridaInternationalUniversity.

13. Hassan,S.,Irfan,A.,Mirza,A.,Siddiqi,I.,2019.CursiveHandwrittenTextRecognitionusingBi DirectionalLSTMs:ACase Study on Urdu Handwriting, in: 2019 International Conference on Deep Learning and Machine Learning in Emerging Applications(Deep ML).IEEE.

14. Hassan,S.U.,Ahamed,J.,Ahmad,K.,2022.Analyticsofmachinelearning basedalgorithmsfortextclassification.Sustainable OperationsandComputers3,238 248.https://doi.org/10.1016/j.susoc.2022.03.001.

15. Joseph, F.J.J., 2019. Effect of supervised learning methodologies in offline handwritten Thai character recognition. InternationalJournalofInformationTechnology12,57 64.https://doi.org/10.1007/s41870 019 00366 y.

16. Joseph,S.,George,J.,2020.HandwrittenCharacterRecognitionofMODIScriptusingConvolutionalNeuralNetworkBased FeatureExtractionMethodandSupportVectorMachineClassifier,in:2020IEEE5thInternationalConferenceonSignaland ImageProcessing(ICSIP).IEEE.

17. Kavitha,B.R.,Srimathi,C.,2022.BenchmarkingonofflineHandwrittenTamilCharacterRecognitionusingconvolutional neural networks. Journal of King Saud University Computer and Information Sciences 34, 1183 1190. https://doi.org/10.1016/j.jksuci.2019.06.004.

18. Kim,C. M.,Hong,E.J.,Chung,K.,Park,R.C.,2020.Line segmentFeatureAnalysisAlgorithmUsingInputDimensionality ReductionforHandwrittenTextRecognition.AppliedSciences10,6904.https://doi.org/10.3390/app10196904.

19. Kim,H.,Lee,W.,Kim,M.,Moon,Y.,Lee,T.,Cho,M.,Mun,D.,2021.Deep learning basedrecognitionofsymbolsandtextsat an industrially applicable level from images of high density piping and instrumentation diagrams. Expert Systems with Applications183,115337.https://doi.org/10.1016/j.eswa.2021.115337.

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page21

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

20. Kumar,P.,Sharma,A.,2020.Segmentation freewriteridentificationbasedonconvolutionalneuralnetwork.Computers &ElectricalEngineering85,106707.https://doi.org/10.1016/j.compeleceng.2020.106707.

21. Li, Z., Teng, N., Jin, M., Lu, H., 2018. Building efficient CNN architecture for offline handwritten Chinese character recognition. International Journal on Document Analysis and Recognition (IJDAR) 21, 233 240. https://doi.org/10.1007/s10032 018 0311 4.

22. Mirza,A.,Zeshan,O.,Atif,M.,Siddiqi,I.,2020.Detectionandrecognitionofcursivetextfromvideoframes.EURASIPJournal onImageandVideoProcessing2020.https://doi.org/10.1186/s13640 020 00523 5.

23.Narang,S.R.,Kumar,M.,Jindal,M.K.,2021.DeepNetDevanagari:adeeplearningmodelforDevanagariancientcharacter recognition.MultimediaToolsandApplications80,20671 20686.https://doi.org/10.1007/s11042 021 10775 6.

24. Naz, S., Umar, A.I., Ahmad, R., Siddiqi, I., Ahmed, S.B., Razzak, M.I., Shafait, F., 2017. Urdu Nastaliq recognition using convolutional recursivedeeplearning.Neurocomputing243,80 87.https://doi.org/10.1016/j.neucom.2017.02.081.

25. Nguyen, H.T., Nguyen, C.T., Ino, T., Indurkhya, B., Nakagawa, M., 2019. Text independent writer identification using convolutionalneuralnetwork.PatternRecognitionLetters121,104 112.https://doi.org/10.1016/j.patrec.2018.07.022.

26. Pande,S.D.,Jadhav,P.P.,Joshi,R.,Sawant,A.D.,Muddebihalkar,V.,Rathod,S.,Gurav,M.N.,Das,S.,2022.Digitizationof handwrittenDevanagaritextusingCNNtransferlearning Abettercustomerservicesupport.NeuroscienceInformatics2, 100016.https://doi.org/10.1016/j.neuri.2021.100016.

27. Ptucha,R.,PetroskiSuch,F.,Pillai,S.,Brockler,F.,Singh,V.,Hutkowski,P.,2019.Intelligentcharacterrecognitionusingfully convolutionalneuralnetworks.PatternRecognition88,604 613.https://doi.org/10.1016/j.patcog.2018.12.017.

28. Qiao,J.,Wang,G.,Li,W.,Chen,M.,2018.AnadaptivedeepQ learningstrategyforhandwrittendigitrecognition.Neural Networks107,61 71.https://doi.org/10.1016/j.neunet.2018.02.010.

29. Rabby,A.S.A.,Haque,S.,Islam,S.,Abujar,S.,Hossain,S.A.,2018.BornoNet:BanglaHandwrittenCharactersRecognition UsingConvolutionalNeuralNetwork.ProcediaComputerScience143,528 535.https://doi.org/10.1016/j.procs.2018.10.426.

30. Rahal, N., Tounsi, M., Hussain, A., Alimi, A.M., 2021. Deep Sparse Auto Encoder Features Learning for Arabic Text Recognition.IEEEAccess9,18569 18584.https://doi.org/10.1109/access.2021.3053618.

31.Shi,B., Bai,X., Yao,C., 2017.AnEnd to End TrainableNeural Network for Image BasedSequenceRecognitionandIts Application to Scene Text Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 39, 2298 2304. https://doi.org/10.1109/tpami.2016.2646371.

32. Shivakumara, P., Sreedhar, R.P., Phan, T.Q., Lu, S., Tan, C.L., 2012. Multioriented Video Scene Text Detection Through BayesianClassificationandBoundaryGrowing.IEEETransactionsonCircuitsandSystemsforVideoTechnology22,1227 1235.https://doi.org/10.1109/tcsvt.2012.2198129.

33. ShobhaRani,N.,Chandan,N.,SajanJain,A.,R.Kiran,H.,2018.Deformedcharacterrecognitionusing convolutional neural networks.InternationalJournalofEngineering&Technology7,1599.https://doi.org/10.14419/ijet.v7i3.14053.

34. Wang, T., Xie, Z., Li, Z., Jin, L., Chen, X., 2019. Radical aggregation network for few shot offline handwritten Chinese characterrecognition.PatternRecognitionLetters125,821 827.https://doi.org/10.1016/j.patrec.2019.08.005.

35. Wang,Y.,Lian,Z.,Tang, Y.,Xiao, J.,2019.Boostingscenecharacter recognition bylearningcanonical forms ofglyphs. International Journal on Document Analysis and Recognition (IJDAR) 22, 209 219.https://doi.org/10.1007/s10032 019 00326 z

36. Wang,Z. R.,Du,J.,2022.Fastwriteradaptationwithstyleextractornetworkforhandwrittentextrecognition.Neural Networks147,42 52.https://doi.org/10.1016/j.neunet.2021.12.002.

37. Wang, Z. R., Du, J., 2021. Joint architecture and knowledge distillation in CNN for Chinese text recognition. Pattern Recognition111,107722.https://doi.org/10.1016/j.patcog.2020.107722.

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page22

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

38. Weldegebriel, H.T., Liu, H., Haq, A.U., Bugingo, E., Zhang, D., 2020. A New Hybrid Convolutional Neural Network and eXtreme Gradient Boosting Classifier for Recognizing Handwritten Ethiopian Characters. IEEE Access 8, 17804 17818. https://doi.org/10.1109/access.2019.2960161.

39. Yan,C.,Xie,H.,Liu,S.,Yin,J.,Zhang,Y.,Dai,Q.,2018.EffectiveUyghurLanguageTextDetectioninComplexBackground Images for Traffic Prompt Identification. IEEE Transactions on Intelligent Transportation Systems 19, 220 229. https://doi.org/10.1109/tits.2017.2749977.

40. Yi FengPan,XinwenHou,Cheng LinLiu,2011.AHybridApproachtoDetectandLocalizeTextsinNaturalSceneImages. IEEETransactionsonImageProcessing20,800 813.https://doi.org/10.1109/tip.2010.2070803.

41. Zhang,C.,Ding,W.,Peng,G.,Fu,F.,Wang,W.,2021.StreetViewTextRecognitionWithDeepLearningforUrbanScene Understanding in Intelligent Transportation Systems. IEEE Transactions on Intelligent Transportation Systems 22,4727 4743.https://doi.org/10.1109/tits.2020.3017632.

42.Zhang,X.,Liu,X.,Sarkodie Gyan,T.,Li,Z.,2021.DevelopmentofacharacterCAPTCHArecognitionsystemforthevisually impairedcommunityusingdeeplearning.MachineVisionandApplication.https://doi.org/10.1007/s00138 020 01160 8.

43.Zhang,Y.,Liang,S.,Nie, S.,Liu,W.,Peng,S.,2018.Robustofflinehandwrittencharacterrecognitionthroughexploring writer independent features under the guidance of printed data. Pattern Recognition Letters 106, 20 26 https://doi.org/10.1016/j.patrec.2018.02.006.

44.Zhao,H.,Liu,H.,2019.MultipleclassifiersfusionandCNNfeatureextractionforhandwrittendigitsrecognition.Granular Computing5,411 418.https://doi.org/10.1007/s41066 019 00158 6.

45. Zuo, L. Q., Sun, H. M., Mao, Q. C., Qi, R., Jia, R. S., 2019. Natural Scene Text Recognition Based on Encoder Decoder Framework.IEEEAccess7,62616 62623.https://doi.org/10.1109/access.2019.2916616

46.Bora,M.B.,Daimary,D.,Amitab,K.,&Kandar,D.(2020).HandwrittenCharacterRecognitionfromImagesusingCNN ECOC. Procedia Computer Science, 167,2403 2409.https://doi.org/10.1016/j.procs.2020.03.293

47.Chaturvedi,S.,Titre,R.N.,Sondhiya,N.R.,Khurshid,A.andDorle,S.,2014.Digitsandaspecialcharacterrecognitionsystem usingannandsnnmodels. International journal of digital image processing, 6(06).

48.Gohil,G.,Teraiya,R.andGoyani,M.,2012.ChaincodeandholisticfeaturesbasedOCRsystemforprinteddevanagariscript usingANNandSVM. International Journal of Artificial Intelligence & Applications, 3(1),p.95.

49.Al Boeridi,O.N.,SyedAhmad,S.M.andKoh,S.P.,2015.Ascalablehybriddecisionsystem(HDS)forRomanwordrecognition usingANNSVM:studycaseonMalaywordrecognition. Neural Computing and Applications, 26(6),pp.1505 1513.

50.Upadhyay,P.,Barman,S.,Bhattacharyya,D.andDixit,M.,2011,June.EnhancedBanglaCharacterRecognitionusingANN.In 2011 International Conference on Communication Systems and Network Technologies (pp.194 197).IEEE.

51.Hassan,S.,Irfan,A.,Mirza,A.,Siddiqi,I.,2019.CursiveHandwrittenTextRecognitionusingBi DirectionalLSTMs:ACase Study on Urdu Handwriting, in: 2019 International Conference on Deep Learning and Machine Learning in Emerging Applications(Deep ML).IEEE.

52.Joseph,S.,George,J.,2020.HandwrittenCharacterRecognitionofMODIScriptusingConvolutionalNeuralNetworkBased FeatureExtractionMethodandSupportVectorMachineClassifier,in:2020IEEE5thInternationalConferenceonSignaland ImageProcessing(ICSIP).IEEE.

53. Panhwar, M.A., Memon, K.A., Abro, A., Zhongliang, D., Khuhro, S.A., Memon, S., 2019. Signboard Detection and Text RecognitionUsingArtificial NeuralNetworks,in:2019IEEE9thInternationalConferenceonElectronicsInformationand EmergencyCommunication(ICEIEC).IEEE.

54.Chowanda,A.,Sutoyo,R.,Meiliana,Tanachutiwat,S.,2021.ExploringText basedEmotionsRecognitionMachineLearning Techniques on Social Media Conversation. Procedia Computer Science 179, 821 828. https://doi.org/10.1016/j.procs.2021.01.099

2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal

International Research Journal of Engineering and Technology (IRJET) e ISSN: 2395 0056

Volume: 09 Issue: 08 | Aug 2022 www.irjet.net p ISSN: 2395 0072

55.Chiong,R.,Budhi,G.S.,Dhakal,S.,Chiong,F.,2021.Atextual basedfeatureapproachfordepressiondetectionusingmachine learning classifiers and social media texts. Computers in Biology and Medicine 135, 104499. https://doi.org/10.1016/j.compbiomed.2021.104499.

56. Wu, C., Fan, W., He, Y., Sun, J., Naoi, S., 2014. Handwritten Character Recognition by Alternately Trained Relaxation ConvolutionalNeuralNetwork,in:201414thInternationalConferenceonFrontiersinHandwritingRecognition.IEEE.

57.Zhang,Y. K.,Zhang,H.,Liu,Y. G.,Yang,Q.,Liu,C. L.,2019.OracleCharacterRecognitionbyNearestNeighborClassification withDeepMetricLearning,in:2019InternationalConferenceonDocumentAnalysisandRecognition(ICDAR).IEEE.

58.Mirza,A.,Zeshan,O.,Atif,M.,Siddiqi,I.,2020.Detectionandrecognitionofcursivetextfromvideoframes.EURASIPJournal onImageandVideoProcessing2020.https://doi.org/10.1186/s13640 020 00523 5

59. Shivakumara, P., Phan, T.Q., Lu, S., Tan, C.L., 2013. Gradient Vector Flow and Grouping Based Method for Arbitrarily OrientedSceneTextDetectioninVideoImages.IEEETransactionsonCircuitsandSystemsforVideoTechnology23,1729 1739.https://doi.org/10.1109/tcsvt.2013.2255396

60. Cilia, N.D., De Stefano, C., Fontanella, F., Scotto di Freca, A., 2019. A ranking based feature selection approach for handwrittencharacterrecognition.PatternRecognitionLetters121,77 86.https://doi.org/10.1016/j.patrec.2018.04.007

61.Abdulrazzaq,M.B.,Saeed,J.N.,2019.AComparisonofThreeClassificationAlgorithmsforHandwrittenDigitRecognition,in: 2019InternationalConferenceonAdvancedScienceandEngineering(ICOASE).IEEE.

2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal