"ARTIFICIALINTELLIGENCEAbstractThe handwritten recognition is generated as one of the most emerging problems in recentdays. Also, it is having high requirement in academics and in commercial sectors. For improvingthe accuracy, the implementation of the handwritten recognition is widely used. This system willbe generating various approaches based on the neutral networks, and these approaches will beused for the recognition of handwritten digits. Then, all the approaches will be compared and thebest approach will be selected in terms of the accuracy. This system is made in terms of readingthe handwritten digits and give approximate end results. In this proposed system, it will presentthe study of different types of approaches in the neutral networks to recognize the handwrittendigits.Keywords: Handwritten Recognition, Neural Network, Types of Approaches, Accuracy.1 Table of ContentsAbstract ........................................................................................................................................................ 11. Introduction ......................................................................................................................................... 32. Methodology ........................................................................................................................................ 33. Simulation Results .............................................................................................................................. 74. Conclusion ......................................................................................................................................... 125. References .......................................................................................................................................... 122 1. IntroductionAn artificial replica of the human brain’s nervous system is a Neutral Network. Anartificial replica signifies that, ‘something is created by human without the help of nature’. Theidentification of alphabets, numerical terms, objects and sounds are the natural habit of humans.Yet, making a machine do these sorts of tasks is an exceptionally confused issue. Handwrittenrecognition is an essential part of computer vision and artificial brilliance.Handwritten digits recognition is an old and critical issue of machine learning. The goalis to perceive image of unique handwritten digits (0-9). More in particular, the issue is identicalto locate a model, which takes handwritten digits image as input, and yields the anticipated classname of the image. Moreover, the thoughts and techniques to take care of this issue would beextremely helpful in different fields of handwritten recognition issues where vast volumes of trueworld-wide information is utilized.Neural system has been generally utilized as a part of the handwritten recognition field.This paper presents comparative studies on handwritten digit recognition using variousapproaches of Neural Network, such as model based on Single layer perceptron (SLP), Back- proliferation (BP) and Hopfield Neural Network (HNN)(Cardoso and Wichert, 2013).2. MethodologyA neural network in artificial intelligence is mainly designed for the process of dataclassification and pattern recognition. Neural network has the process that is similar to theprocess of human brain. This network provides the application of handwritten digit recognition.This handwritten digit recognition system has the following phases(ca and Gupta, 2015).? Acquisition ? Image Preprocessing? Segmentation ? Image Classification? Feature ExtractionThe first phase of the process is image acquisition. This method is used to obtain theinput image by capturing photograph or by scanning the documents. In the second phase,preprocessing is done on an input image. Preprocessing means applying such techniques on animage such as smoothing, normalizing, filtering, thresholding and resizing. After that, in these3 processes the image is finally applied by using the classification technique. The next phase issegmentation of an image. It results in splitting of the image into various segments. It leads toanalyzing the image into more meaningful manner. The next phase is feature reduction. Itcaptures the main characteristics of an image and is related to the reduction in dimension. Theoutput of this stage gives an input image to the classification technique. Neural Network (NN) isa best classifier to classify an image(ca and Gupta, 2015).There are various classification approaches that are available in the neural networks.They are,? Single layer perceptron model? Hopfield Neural Network? Back-propagation AlgorithmThe unknown pattern fall area will be detected by the classification method. Thisprovides the extraction of exact region of feature space. Back-propagation algorithm is chosenfor this handwritten recognition process. As compared with the other approaches, this algorithmgives minimum error rate at the level of network weights. In the single layer perceptron modelapproach, the network is consider as a single neuron. This will take different input signals and4 provides a single output signal. It gives an output signal based on the weighted input signals.This approach is limited in the combined network and also in network which uses differentactivation functions. In the Hopfield neural network approach, this can be performed in thecombined network. An output of one network is given to an input of the other networks or units.It uses the pattern storage task. This task is used to store the set of patterns and matches thepattern that are given to the network. The features of an image and their spatial relationship isstored in the pattern storage task. This approach contains some limitations such as hard problemand false energy minima problem(Qureshi et al. 2011).Single Layer Perceptron ModelThe single layer perceptron model’s simple construction is generated in the below figure.The process of the input signal delivering the output signal will be considered as the term singleneutron.The output value will be based on the relative strength of the considered input signals.The output of the single layer perceptron will be generated as, The adaptive weight vector will be considered as the w(n)=[w (n)…………w (n)]. The1 N input signal vector will be considered as x(n)=[x (n)…………x (n)]Tand the input signal vector1 N will be denoted as T. The bias term will be termed as b. The frequently used functions foractivations are the sigmoid and the hard limiter.The updated perceptron weights will be basedon the below equations,5 The learning rate parameter will be denoted by ? and it well be less than 1. The desiredoutput is denoted by d(n)(Cecotti, 2016).Hopfield Neural NetworkIn this approach, the output of each unit is forwarded to the input of the other units. Ituses the pattern storage task. This task is used to store the set of patterns and matches the patternthat are given to the network. The features of an image and their spatial relationship is stored inthe pattern storage task. The above image is the simple Hopfield network. For storing the pattern,the algorithm is used for Hopfield network. The algorithm is given below:Consider the vectors that are stored is a1, 1=1, 2.... L. Step 1: Initialize the output pattern with the input that is given as the unknown pattern. S (0) = a , where i = 1, 2,....Ni i Here, S (0) is the output generating function at time 0.i Step 2: Assign weights for each connection with each unit.fori ? j= 0for i = j,1 = i , j = NStep 3: Iterate the above process until it obtains the convergence.fori = 1,2,3……..,NThe output of the algorithm provides the best matches based on the input. This is calledas the stored pattern(KSarangi, K Sahoo and Ahmed, 2012).6 Back-propagation AlgorithmBack propagation algorithm contains two phases. They are forward phase and thebackward phase. The neural network consists of three layers. They are input layer, hidden layerand output layer. The forward phase of the back-propagation algorithm moves the activationsfrom the input layer of the network to the output layer. In the second phase (backward phase),the requested nominal value and the actual value that is observed are moved to the output layer.So, the bias value and the weights are changed. The algorithm for the back-propagation isdescribed below(Shinde&Waghulade 2016).Step 1: First initialize the input layer parameters. They are W = random weights, I =i i input, T = target values and Y = output.i i Step 2: Forward the activations from input layer to the output layer.I => H => OStep 3: In each layer, the output of each neurons is calculated.O =? I * W i i iStep 4: Final output is calculated from each neurons when apply activation functions on-(0)each neurons.Y = 1/1+e i Step 5: Output layer error is calculated using below formula.Error = 2(Y –T )i i Step 6: Propagate the error backwards.dE/dW = d(Error)/dW t tStep 7: Weights are updated using below formula.?W = - e (dE/dW ) + a( ?W )t t t-1Here, learning rate is e and momentum is a.3. Simulation ResultsVarious data sets are used for the image of handwritten recognition. The commonly useddata sets are USPS, MNIST and DIGITS. MNIST and USPS data sets are widely used for theassessment of handwritten digit classification and recognition. These data sets are classified intotest and training data set. Classifier process uses test data set. Preprocessing, Feature extraction,segmentation process uses training data set for obtaining the results.The combined data sets7 provides the increased recognition accuracy. The vector e-grc3 produces high performance bycomparing with e-blc3. This vector produces low error rate (Cecotti, 2016). Figure 1 : Vector Error Rate The neural network was developed in MATLAB version 2012a. It also consists ofPentium IV system with 3.4 GHz and 4GB Memory. The scalar output range is between (0, 1).All the neutrons contain same ranges. The Weka tool is good. It uses Random Forest algorithm.In machine learning, the Naïve Bayes Multinomial is used. But, the major problem in Wekasoftware is that, it does not have the capability of testing the data classification. For obtainingaccurate results, the handwritten image recognition uses 'k cross fold technique'. The programused in this system does not execute in any other applications. The picture shows the results forusage of various algorithms such as Lazy.IBI, Rules.NNge, Rules. PART, Naïve BayesMultinomial and Back propagation. Among these algorithms,only the Backpropagation produceshigh accuracy of results. This is the main reason for choosing Back propagation algorithm in thehandwritten image recognition performed in MATLAB 2012a(N and B, 2016).8 Figure 2 : Accuracy of Various AlgorithmsThe Backpropagation Algorithm produces less mean square error while increasing theiterations. But, the mean square error is constant after a particular point. By comparing with theother machine learning in the neural networks, BPA takes more time. The experiment iscompared with other image techniques for the similar data set. But, the algorithm is a failure.Thus, BPA algorithm is the best one for the handwritten digit image recognition. Because, it useseach feature vectors present in the numerical form (Automatic Recognition of HandwrittenDigits Using Multi-Layer Sigmoid Neural Network, 2016). The backend used for the process isGUI. Figure 3 : Handwritten Digit Image? GUI receives the hand written digit as query image.? The noiseless picture appears for the loaded data.? Extraction of data will be performed.9 ? Final output will be obtained for the corresponding input image. The simulation result for the tested data and the trained data is given below,Figure 4 : Result of Training data Figure 5 : Result of Test DataThis handwritten digit images are categorized into histograms. The neural networkacquires these histograms for executing the process. This handwritten digit images are finallyconverted into binary digits in 0’s and 1’s. The result will produce the score of output based onthe input digit. The input is possible between 0-9 digits. The handwritten digit is a simple image,but contains 100 samples. The accuracy is about 25 times to meet the result to 99%.The result of handwritten digit input is processed using the image editor. The trainingdata set is tested using the paint brush which is an image editor tool. This simulation results canbe used for Signature Pattern matching, Handwritten AlphabetRecognitionetc.(RomeshLaishram, 2015). This system can be implemented for the otherlanguages such as Malayalam, and Urdu characters.10 In the single layer perceptron neural network, the error rate is 9.98% for the bitmapformat 32X32. In the Hopfield neural network, the error rate is more when compared with the otherapproaches.11 4. ConclusionThis paper presents the handwritten recognition approaches using the neural networks. Inthe recognition process, the back-propagation algorithm is the most successful approach. It is thefast and efficient approach when compared with other approaches. This approach gives thereduction in computing time. Also, it provides simple implementation. If the selected weight issmall at the beginning of the process, then it will give small amount of computation time. Wecan infer that we have achieved the computer to thehuman's brain by the significance utilizationof unique digits recognition for various applications. This recognition begins with obtaining theimage to be inclined in various steps. As an essential point, categorization and acknowledgmentmust be done to pick up a numeral content. At last, neural network is best and itis better than anyother techniques of the handwritten recognition.5. ReferencesAutomatic Recognition of Handwritten Digits Using Multi-Layer Sigmoid Neural Network.(2016). International Journal of Science and Research (IJSR), 5(3), pp.951-955.ca, S. and Gupta, D. (2015). Handwritten Digit Recognition using various Neural NetworkApproaches. IJARCCE, pp.78-80.Cardoso, Â. and Wichert, A. (2013). Handwritten digit recognition using biologically inspiredfeatures. Neurocomputing, 99, pp.575-580.12 Cecotti, H. (2016). Active graph based semi-supervised learning using image matching:Application to handwritten digit recognition. Pattern Recognition Letters, 73, pp.76-82.KSarangi, P., K Sahoo, A. and Ahmed, P. (2012). Recognition of Isolated Handwritten OriyaNumerals using Hopfield Neural Network. International Journal of Computer Applications,40(8), pp.36-42.Lotfi, A. and Benyettou, A. (2011). Using Probabilistic Neural Networks for Handwritten DigitRecognition. Journal of Artificial Intelligence, 4(4), pp.288-294.N, V. and B, R. (2016). Segmentation and Recognition of Handwritten Digit Numeral StringUsing a Multi Layer Perceptron Neural Networks. International Journal in Foundations ofComputer Science & Technology, 6(1), pp.49-55.Qureshi, M., Qureshi, M., Nabi, M. and Khalid, S. (2011). Handwritten digit recognition systemusing neural network. Energy Procedia, 13, pp.4326-4336.RomeshLaishram, (2015). Simulation and Modeling of Handwritten Meitei Mayek Digits usingNeural Network Approach.Saeed AL-Mansoori (2015). Intelligent Handwritten Digit Recognition using Artificial NeuralNetwork, 5(5), pp.46-51.Shinde, S. and Waghulade, R. (2016). Handwritten Mathematical Expressions Recognition usingBack Propagation Artificial Neural Network. Communications on Applied Electronics, 4(7),pp.1-6.Viragkumar, N. (2014). Fast Efficient Artificial Neural Network for Handwritten DigitRecognition, 5(2), pp.2302-2306.13 "