Classification System for Handwritten Devnagari Numeral with a Neural Network Approach

This paper addresses an important and vital problem within the general area of character recognition, namely recognizing Marathi handwritten numerals. Artificial neural network approaches have been recognized as a powerful tool for handwritten numeral recognition. This paper demonstrates the use of single hidden layer MLP NN as a classifier for handwritten Marathi Numerals of Devnagari script. In present study, a MLP NN is designed with Tan sigmoid activation function for hidden and Log sigmoid function for output layer with neurons in hidden layer varied from 16 to 128 in steps of 16, constitutes 8 configurations of MLP NN trained three times each with memory efficient and fast Scaled Conjugate Gradient (SCG) algorithm. An image (64x64) of handwritten digits act as an input to the network, the training is controlled by early stopping criteria so that optimal network is derived. The intended network is analysed on various performances metric such as mse, best linear fit, correlation coefficient and misclassification rate. The scruples analysis of the result on different data partitions such as training, validation and testing provides best network to be further analysed. Further it is shown that the average classification accuracy for the best network is 98.35%, 89.71%, 91.28% and 96.77% on training, validation, testing and overall dataset respectively. On the basis of confusion matrix, results are elaborated with % misclassification for each output class distributed uniformly within dataset of 4465 samples. Network complexity in terms of weights and bias is 492938 connections from input to output. Keywords— Handwritten Numerals recognition, MLP, Scaled Conjugate Gradient (SCG) algorithm, best regression fit, Confusion Matrix, log-sigmoid, tan-sigmoid.