Generative Adversarial Image Generation for Handwriting Recognition

Abstract

The problem of handwritten text recognition is widely considered a solved problem, both in-n-out of the engineering community. However, these models perform better only in one of two extremely confined conditions: recognizing printed text and recognizing texts which were fed to it at the time of training. It is still a challenge to develop a robust state of the art model that performs well on a wide variety of free handwritten texts. The lack of large annotated datasets is one of the primary reasons. Gathering data is a costly business and annotating the data makes it even more difficult. We present method of synthesizing handwritten text images with different writing styles, using GANs. Handwriting recognition on Indian regional scripts is less explored and it is a very interesting and potential research area. We used DCGAN on Bangla Script to generate fake images which will help to increase data and bridge between supervised and unsupervised learning.