OnkoGan: Bangla Handwritten Digit Generation with Deep Convolutional Generative Adversarial Networks

Abstract

From a very early age human achieve a precious skill that is a handwriting. After this invention, the ardor of it changed day by day. And every human has a different style of handwriting. So, facsimile anyone’s handwriting is a difficult task and it needs the strong ability of brain and practice. This paper is about this mimicry where an artificial system will do this by using Generative Adversarial Networks (GANs) [1]. GANs used in unsupervised machine learning that is implemented by two neural networks. GANs has a generator which generates fake images and a discriminator which make a difference between a real image and a fake image. We trained our proposed DCGAN [2] (Deep convolutional generative adversarial networks) to achieve our goal by using the three most popular Bangla handwritten datasets CMATERdb [3], BanglaLekha-Isolated [4], ISI [5] and our own dataset Ekush [6]. The proposed DCGAN successfully generate Bangla digits which makes it a robust model to generate Bangla handwritten digits from random noise. All code and datasets are freely available on https://github.com/SadekaHaque/BanglaGan.