Currency Detection using TensorFlow
Shelar Rutuja1 , More Smruti1 , Tapase Nisha1 , Sanjay Waykar1
1MGM College of Engineering and Technology, Kamothe
:Advancesintechnologyhavereplacedpeoplein almosteveryfieldwithmachines.Thankstotheintroduction ofmachines,bankingautomationhasreducedtheburdenon humans. Banking automation requires more attention to decliningcurrencyhandling.Whenthebanknoteisblurredor defaced, it is difficult to identify its currency value. A sophisticateddesignisincludedtoincreasethesecurityofthe call.This makesthecall recognitiontaskverydifficult.For correctcurrencyrecognition,itisveryimportanttochoosea goodfunctionandanappropriatealgorithm.Oneofthemain problemsthatblindpeoplefaceistherecognitionofmoney, especiallycash.Inaway,theseeminglyweakenedpeopledo not think about cash settlement and run into problems related tocashtransactionsintheirdailylife.Itisa useful treatment for those who are externally weakened. studies and trialswereconductedaccordingtokeypoints,suchas watermarks, imagesprintedonmoney,thevalueofwords andnumbers,andthetotalamountofinformationgathering that stimulated CNN .This paper focuses on the study of solvingsocialproblemsusingConvolutionalNeuralNetworks (CNNs)andvalidatingandevaluatingdifferentCNNmodels. Here, the Alexnet, Googlenet and Vgg16 models were considered for the study .All models were adjusted during preparationandtestingofindividualdatasets.Amongthese three models, Alexnet had the best performance, Vgg16 modelshowed100%performance,andGooglenetshowed performancewith88%.
Keywords banknoterecognition;convolutionalneural networks; computer vision;deep learning;VGG16; transfer learning
I. INTRODUCTION
Currencyisnotesandcoinsissuedbythegovernmentfor circulationintheeconomy.Serviceandproduct exchange facility. Banknotes are an important medium for trading. The banknote is characterized by simplicity, durability,fullcontrolandaffordability.Thismadehimwell known.Comparedtoallotheralternativeformsofcurrency, paperisthemostpreferredformofcurrency.Thishasone downsidetobanknotes:itcannotbereused,buttheissueis less serious compared to other methods As part of the technological advancements introduced to the financial sector, financial institutions and banks have embarked on financial self service. An automated banking system is implemented that processes currencies using, machines withATMcountersandcoindispensers.Inthissituation,the usesacurrencyrecognizertoclassifythebanknotes.Callhas
two types of features i.e. internal features and external features.Externalcharacteristicsincludephysicalaspectsof thecurrency,suchaswidthandsize.However,thesephysical features are unreliable as circulation can damage the currency.Thiscompromisedcallsystemcancausethesystem tonotrecognizethecall.Internalfeaturesincludeunreliable colorfeaturesbecausecallsgothroughdifferenthandsand thiscangetdirtyandgiveincorrectresults.Forcurrenciesof eachdenomination there is a specific colorand size thatis followed by Reserve bank of India. It is a very simple for human to identify them denomination of currency note becauseourhumanbrainisextremelyskillfulinlearningnew matters and discovering them later without much trouble. But this currency recognition task turns very difficult and challenging in computer vision, in cases whencurrencies becomes damaged, old, and faded due to wear and tear of currencies. Security features are included in every Indian Currency which dispenses help in recognition and identificationofthecurrencyvalue.VariousSecurityfeatures areidentificationmark(shape),Centervalue,Ashoka,Latent image, See through register, Security thread, Micro letter, WatermarkandRBIseal.
1.1 Deep Learning: Deep Learning is a new machine learningenvironment,introducedwiththeaimofbringing Machine Learning closer to one of its original motive: ArtificialIntelligence”(p.6).In depthreadingreferstothe study of multiple categories of illustrations and concepts thatassistresearchersinanalyzingdata,includingimages, sounds,andtexts
In depth reading is often associated with a neural networkwithafewlayersthatcanlearnfromalargeamount ofdata,whichincludesaseriesoflabeledimages.Inaddition it has been widely used in the field of vision and voice (processingnaturallanguage).Theweights ofanylayerare learned through the spread of back propagation in an in depth learning process.All layershave different effects on data analysis. Despite its difficulties, the method has been successfully extended to a variety of classification and identificationproblems.
1.2 Deep learning techniques:
Classic Neural Networks: A multilayer perceptron, inwhichneuronsareassociatedwithacontinuouslayer,and isoftenusedtoidentifyNeuralConnectedNetworks.There are3functionsinvolvedinthelaunch system:Linefunction: Properly called, this indicates only the line that repeats the input with a continuous repetition. Non line work: Thisiscategorizedintothreecategories:A “SigmoidCurve: A function that is translatedas anS shaped curvewith its widthstartingfrom0to1.
B Hyperbolic tangent (tanh)
to an S shaped curvewithadistanceof1to1.
C Rectified Linear Unit (ReLU): A single point function thatreturns0whentheinputvalueisbelowthesetvalueand the line multiplication if the input value is high. Convolutional Neural Networks: CNN is a continuous version of the highly developed neural network model model.Itisdesignedtohandlehighlevelsofcomplexity,as wellaspre processinganddataintegration.
CNNisdividedintofourstagesaftertheinputdatawas presentedintheconversionprogram:Convolution:aprocess thatgeneratesfeaturemapsfrominputdatabasesanduses functioninthesemaps.
Max Pooling: Assists C NN in image acquisition based onimaginarymodification.
Plans: At this stage, the data that has been generated is thenpressedforCNNtocheck.
CompleteConnection:Thisissometimesdescribed as ahiddenlayerthatcollectssystem lossfunction.
Recurrent Neural Networks (RNNs):initially,RNNwas developedtoassistinsequencingprediction;forexample,the Long Short Term Memory (LSTM) algorithm is widely knownforitsversatility.Thesenetworksarebasedsolelyon datatrackingwithvariableinputlengths.
Generative Adversarial Networks: Using an efficient production system for counterfeit production models and detectingreal timedatafromproducers,knownas(GANs).In their research, they used GANs and could distinguish between fake and real banknotes with a high degree of accuracy. GAN.sis a two dimensional and highly complex neuralnetwork.Thegenerative networkisthefirst,andthe racist network is the second. After using GANs in the database and using them to separate real and fake notes, promisingresultswereobtained.
1. Single shot multibox detector: It is a comprehensive mathematicalmodelinfindingobjects.Itcreatesconnecting boxes using working maps from a variety of layers. It will builddifferentboundaryboxesbasedondifferentcategories; afterthat,onecanfindoutwhattheobjectwas.Thissystem can also be used for real time detection, and has faster capabilitiesthanR_C NN and Res Net.
2. Multilayer perception:Thisis oneofa kindofneural network built into a layer of input and a node that works togethertoproduceoneormorehiddenlayersandeffects. Tocalculateoutput,indirectactivityisrequired,andaback propagation algorithmcanbeusedtotrainMLPisolation andreduction.Inaddition,MLPswereafeedforwardneural network composed ofa number of nodes connected by a singleconnectionandtrainedusingback propagation.
II.EXISTING SYSTEM
Theauthorsof[1]proposedamethodofrecognizingmoney using monetary features such as central number, latent image,RBI,form,andmicroletter.Itcontainsatrainingdata setforpreparingthetrainingmodel.PCAanalysisaccounts for banknote recognition. In [2], the authors proposed a method for recognizing banknotes using features such as color, texture, and size. Dirty banknote detection method activated. Author [3] proposed a banknote recognition systemusingMATLABtools.PCAanalysisaccordingtowith a description of Euclidean distance. Describes the LBP method for matching purposes In [4], the authors first determine the country and then propose a method for determiningthecountryname.Theauthorof[5]gaveabrief introduction to the features of the Indian currency. Shape filtering process according to with other analysis and segmentationdescriptions.Abanknotedetectionrecognition technologyisimplementedusinganeuralnetwork [6]isa bookonCNNswhichincludesvariousprocessingoflayers with mathematical demonstrations. Once an image is captured,itistransformedtomatrixformpriortodetection andthenmatchedwithatrainedmodeltodemonstrateall processingbetweenmatrixgenerationandrecognitionusing mathematical calculation. Effect of various convolution kernelsonhorizontalandverticaledges.In[7],theauthors describedallaspectsofCNNs.In[8],theauthorproposeda method for recognizing currency using a neural network. Classificationusingweighted
Euclidean distance with various steps required for data collection and processing. In [9], [10], the authors proposed a call recognition technique, and is a learning machineusedtotraindataandobtainaccuracy
METHODOLOGY
1:Methodologyofthesystem
thecodingsection,wewillusethein depthreadingof Keras library on python to build our CNN (Convolutional Neural Network). We have to include TensorFlow and TheanoworkingattheendoftheCameras.TensorFlowisan organization.Anopensourcesoftwarefordataflowsystem across the entire job list. It is a symbolic mathematical library,anditisandisusedinmachinelearningapplications
suchasneural networks.TheanoisaPythonlibraryandis doingwellmathematiciantoconvertandevaluatestatistics expressions,especiallythosewithamatrixvalue.InTheano, thefiguresareexpressedusingaNumPy squaresyntaxwas alsocompiledforoptimalperformanceitcanbeCPUorGPU architecture.Afterinstallingtherequiredlibraries,wetrain our model asdiscussed above. After training and model testing,wesettheamountoftimethatincreasestheaccuracy ofAFCRSinadditiontoincreasingthenumberofperiods
Data pre processing Aneasywaytogetdatawithout overloadingonceunderthepre processingofthedataset. Themainpurposeafterpre processingthedatathatadding valuebasevaluewhichisasetofdatagenerated.Mainthe advantageofpre processingdataisthatitisbettertraining set. For these purposes, we use the Keras library to pre process images. The VGG 16 model we use is ours experimentsrequireinputimageformat64x64x3,where3 referstoR,G,B(red,green,blue)partsofthecoloredpicture andtheimageshouldbe 64x64pixelsinsize.Wethenuse the following three types of pre image processing in the originaldatabases,tooselectourfirstfilterthatstartsat64.
a. Image Rescaling We need to resize the image to makethemodeldataintoacommonformatfortrainingto bedeveloped,accurate,andfast.Sinere scalingfeatureon thecameras.Tousethisfeature,weneedtoimportacertain libraryfromkeraspre processingas"ImageDataGenerator".
If the re rating feature is missing or 0, no resume is used, otherwiseweduplicatethedataaboutitgivennumber.This isdoneafterinstallingalltheotherschanges.Inourmodel, weusere scalingfeaturelike:
rescale=1./255inbothtrainingandtestingdatasets.
b. Image Shearing Weneedtocuttheimagetomake thetrainingdatabetter,moreaccurate.Wealsohaveshear rangefactorinthecameras.Thiswillbeincorporatedinto thecamera'spre processinglibrary.Inourmodel,weusea clipboard asshear_range = 0.2. Shear range is usually the Shear angle in the middle anti clockwise direction by degrees,alsoknownastheshearintensity.
c. Perspective Transformations: Appliedperspective transformations on training data to zoom in the range of zoom_range=0.2togettheaccurateresultsbylearninginan accuratemanner.Zoomrangeisafloatorlower,upperrange forrandom zoom. This is also done by importing a library fromkeraspre processing.
2. Training the CNN: Here,afterchoosingVGGNetfor ourmodel,weoptimizedVGGNet[4],whichisapre trained network.Thisspeedsupthetrainingprocess,astheyarefew andfarbetweenreallytraininglayers.Trainingtheneural network, of course it is actually better to start with a malfunctioning neural network and expose the neural networkwithhighaccuracy.Astermsofjobloss,wewant ourjoblosstobethesameverylowattheendoftraining. Thisshowsthat
The neural network has a high level of learning and accuracy. The problem of network training equals productivityaslossfunctionwithasmallerrorrate.Itisalso
importanteffectiveevenreduceslossesbecause,itturnsout lossisaneasytasktoimprove.Eventhoughtherearemany algorithmsthatmakeitworkandforbetterperformance,we selectReLU (linearRectifierunit)asouractivationfunction, andwecan select'adam'asourdevelopmentworkbecause, it develops neural networks with faster training. And also computational
TheReLUstepissimple.Toseeanoteofagivencurrency asfakeorreal,weshouldseetheaccuracyofVGG 16awell configuredmodelwithasetofgenerateddata.Itwasabout 55%onthecorrespondingtestset,thoughoursthedataset wasverysmallandlimitedto200images,theresultisstill veryencouraging.Whenweenlargeourimagedatasetwith real world samples that can make the model even higher accurately trained which may result in our results being higher80%accuracyisagoodindicatorofresults.Without fullpoints,theresultmaybeconsideredtoinvestheavilyin the database,a was thetraining andtestsetsverysimilar. Therefore,aftercarefulconsideration,wemustevaluatethe lossandaccuracyofourmodel,bychangingthesizeofthe collection and epoch. Basically there are 2 cases in deep learningaboutlossandaccuracyprices.
Figure2.Testingprocedureofthemodel
IV. RESULTS
The three layer CNN model used to classify monetary notesbasedonsystemshasyielded98.50%accurateresults and 15 times. The training database is divided byan8:2 ratingforoppositevalidation.Multi measurementtemplates areusedtodeterminesecurityfeaturesinacurrencynote.To testthemodel,ascannedimageofthecurrencynotewillbe uploadedviatheJupyterFileUploadSystemanditsvariability is predicted by model. The use of CNN has several advantages,whichinclude,forexample,CNNiswellknown for its architecture, and the best part is that no feature is required.ThegreatadvantageofC NNoveritspredecessoris thatitcanidentifyanimportantfactorwithouttheneedto interactwithpeople.
SomeofCNN'sobstaclesare,Imagesindifferentpositions are separated; due to operations such as max pool, the Convolutional neural network is very slow, If CNN has multiple layers, the training phase can take longer, if the machinedoesnothaveapowerfulGPUandCNNrequiresa largeData settoprocesstrainedneuralnetwork
Figure3.AccuracyofCNN
V. CONCLUSION
Today,technologyisadvancingonamassivescale.The proposedsystemcanbeextendedtodetectcoinsaswellas detect counterfeit currency. You can add the names of countries other thanIndia,youcanalso compare between them.Itdoesnotgive100%accuracywhenimageisloaded externallyintopracticefolder.Youcansolvethisproblemby optimizing the system. Therefore, the various methods proposedinthisarticleweresuccessfullyimplementedand validatedbyrunning experimentsonthemodel.CNNturns outtobethebestfeaturetodothistechniqueusingtfliteand flutter_tts modules. 95% accuracy is achieved with model classifications. Also, detection of coins works well in this method.
REFERENCES
[1]Vishnu R, Bini Omman, "Principal Features for Indian CurrencyRecognition",2014AnnualIEEEIndiaConference (INDICON), Pune, India.doi:10.1109/INDICON.2014.7030679
[2]H.Hassanpour,A.Yaseri,G.Ardeshiri,"FeatureExtraction for Paper currency Recognition", 2007 IEEE 9th international Symposium on signal processing and its Applications,doi:10.1109/isspa.2007.4555366
[3]KalpanaGautam,“IndianCurrencyDetectionUsingImage RecognitionTechnique",2020IEEEInternationalConference on Computer Science, Engineering and Applications(ICCSEA),doi:10.1109/ICCSEA49143.2020.9132 955
[4]Vedasamhitha Abburu, Saumya Gupta, S. R. Rimitha, Manjunath Mulimani, Shashidhar G. Koolagudi, “Currency Recognition System Using Image Processing",2017 IEEE Tenth International Conference on Contemporary Computing,doi:10.1109/IC3.2017.8284300
[5]Ms.RumiGhosh,Mr.RakeshKhare,“AStudyonDiverse Recognition Techniques for Indian Currency Note", 2013(JUNE)INTERNATIONALJOURNALOFENGINEERING SCIENCES&RESEARCHTECHNOLOGY,ISSN:2277 965
[6]Jianxin Wu, “Introduction to Convolutional Neural Networks”,May1,2017.
[7]Simon Haykin, "Neural Networks: A comprehensive foundation",2ndEdition,PrenticeHall,1998
[8]Muhammad Sarfraza, “An intelligent paper currency recognition system", 2015, International Conference on Communication,ManagementandInformationTechnology (ICCMIT2015).
[9]Faisal Saeed, Tawfik Al Hadhrami, Fathey Mohammed, Errais Mohammed,” Advances on Smart and Soft Computing".2021,Springer.
[10]Minal Gour , Kunal Gajbhiye, Bhagyashree Kumbhare, andM.M.Sharm,“PaperCurrencyRecognitionSystemUsing Characteristics Extraction and Negatively Correlated NN Ensemble", 2011, Advanced Materials Research Vols 403 408(2012)pp915 919©(2012)TransTechPublications, Switzerland,doi:10.4028.
[11]TeachableMachine,“https://teachablemachine.with google.com/train/image".
[12]Rubeena Mirza1,Vinti Nanda2 “Design and implementation of Indian Paper currency authentication system based on Feature extraction by Edge based Segmentationusingsobeloperator”Internationaljournalof engineering Research and development e ISSN:2278 067X,p ISSN:2278 800X, www.ijerd.com Volume 3,Issue 2(august2012)PP.41 46
[13]Rubeena Mirza, Vinti Nanda, “paper currency verification System Based On Characteristics Extraction Using Image Processing”, International Journal of engineering and advanced technology (iJEAT) ISSN 2249 8958,Volume 1,Issue 3,February2012.
[14]Kalyan Kumar Debnath, bSultan Uddin Ahmed,aMd. Shahjahan “A Paper currency Recognition System using negativelycorrelatedNeuralnetworkEnsemble”,JOURNAL OFMULTIMEDIA,VOL.5,NO.6,DECEMBER2010@2010
[15]Megha Thakur, Amrit Kaur, “VARIOUS COUNTERFEIT CURRENCY DETECTION AND CLASSIFICATION TECHNIQUES” International Journal For Technological ResearchInEngineeringVolume1,Issue11,pp.1309 1313, July 2014