Evaluating Generative Adversarial Networks for Gurumukhi Handwritten Character Recognition (CR)

Abstract

With the invention of deep learning, optical CR (OCR) has achieved state-of-the-art performance for Latin, Arabic and Chinese scripts. However for regional languages, lack of large and high-quality training data needed to train the deep neural networks is a major barrier for the usage of deep learning. To produce a diverse synthetic data similar to realistic data, a class of deep networks called Generative Adversarial Networks (GANs) were proposed. GANs have performed efficiently for non Indic scripts and in various other fields like object detection, image translation, etc. Gurumukhi script is one of the regional Indic scripts widely used in North India and some other parts of the world like Canada, USA, etc. In this work, we have evaluated the power of GANs for generating the character dataset of regional language of India. Experimental work has been done using variants of GANs to generate the synthetic data for Gurumukhi script. Deep convolution GAN (DCGAN), Wasserstein GAN (WGAN), Least square GAN (LSGAN) and Conditional GAN (CGAN) have been evaluated for the Gurumukhi handwritten characters dataset. GANs with different network structures, training methods, loss functions and optimizers are used and compared for Gurumukhi handwritten character dataset to find the optimal parameters and results.