Isolated Offline Tamil Handwritten Character Recognition Using Deep Convolutional Neural Network

Abstract

In this paper, a Convolutional Neural Network architecture (ConvNet) for offline isolated Tamil character recognition is proposed. A first-ever attempt has been made to recognize all 247 characters in the Tamil text using 124 unique symbols. The proposed architecture contains two Convolutional layers and Two Fully Connected (FC) Layers with ReLu activation function. Softmax function is used in the final layer to compute the probability of the classes. The 9.6 million parameters of the network are randomly initialized using He initialization and fine-tuned using Nesterov Accelerated batch gradient descent optimization algorithm. Dropout regularization method has been used to avoid over-fitting of the network to the training data. A total of 98,992 image samples from IWFHR database are divided into 69% for training set (68,488), 20% for validation set (20,584) and 11% for test set (9920). Cross entropy loss has been used during the training phase to measure the loss and thereby update the parameters of the network. The network has achieved 88.2% training accuracy and 71.1% testing accuracy. The reason for reduction in the test accuracy is analysed. The source code and the dataset have been published for a quicker reproducibility of the result.