OCR-friendly Image Synthesis using Generative Adversarial Networks

Abstract

With quality being a deciding factor in the accuracy of OCR predictions for a given source image, there comes the need of pre-processing methods to improve the quality of an image before undergoing this process. Towards that, we present a GAN-based method targeted at improving the quality of source image in the fields of image resolution, blur and noise. The model uses an encoder trained to generate latent image representations for a low-quality image, the representations corresponding to blur and noise types present in the image. These representations act as inputs to the constructed conditional GAN. Besides these labels, the generator takes a low-quality image as input and is trained to generate a high-quality image as represented by the target images.