HANDWRITTEN DEVANAGARI CHARACTER RECOGNITION USING HYBRID CONVOLUTIONAL NEURAL NETWORK

Abstract

Handwritten character recognition is very popular field of research work in modern technology as information are stored and used as written script in different languages since long ago. This research work mainly focuses to recognize handwritten Devanagari character through preprocessing the image data and deep learning techniques. The handwritten character is collected and processed to fit into popularly used neural networks for classification of those images into their related label of character. Collected image data are clipped, normalized and processed and then converted into numpy dataset to feed into the neural network that we adopted and developed. The data is trained through the managed sequence of convolutional and fully connected layers of network and proper activation and pooling is done in between to optimize and speed up the training process. Here we have used ReLU as activation function and maxpooling as pooling function. We have used two convolution layer and each Convolution Layer are followed by activation and pooling functions. These layers are then followed by three fully connected layer to produce better neural network. 12051 images of handwritten Devanagari character are fed into this neural network and the trained to produce a neural network and the accuracy of this produced neural network is tested upon the different set of 290 validation images test set. The validation of this model is observed through the confusion matrix and seems to work good.