PERFORMANCE ANALYSIS OF DATA AUGMENTATION AND GAN IN DEEP NEURAL NETWORKS FOR HANDWRITTEN TAMIL CHARACTERS
In recent years, significant achievements have been made in the field of handwritten character recognition using Convolution neural networks. Even though large amounts of training data are required to obtain satisfactory results, there is limited number of data only available for training and validation. Changing the size and position of the images, Rotating the images one degree at a time using power point, adding white noise to each image using OpenCV, data pre-processing in, and incorporating bootstrap code in MATLAB are the possible ways of increasing images. However, it is difficult to generate a greater number of images using the above techniques. Also, the quality of the images is not up to the level. Hence to increase the size of the dataset without degrading the quality, we incorporate the techniques called Data Augmentation and Generative Adversarial Networks in recognizing handwritten characters. Data augmentation and GAN are the process of generating new samples and it introduces the changes in the training images for better performance.It is an attempt to analyze the performances ofAugmentation and GAN on the dataset HPLabs. It is found that Augmentation with CNN produced the accuracy of 93% whereas, hybridized GAN with CNN yield the accuracy of 97%.