An ensemble English Document Image Dewarping using Retinex and Generative Adversarial Neural Networks

Abstract

Document images obtained from bound or folded documents mainly undergo a specific image degradation known as document warping, making it difficult to read and significantly diminish the performance of an Optical Character Recognition (OCR) system. This paper attempts to develop a comprehensive framework for dewarping and restoring highly distorted document images using the Retinex theory and Generative Adversarial Networks (GANs). To the best of our knowledge, we could not find much research conducted on this topic using Retinex and GANs for ensembling. It is observed that the proposed method can perform better in tasks, like illumination correction, binarization, edge detection, and dewarping in order to reconstruct an improved version of the distorted text. Further, to demonstrate efficacy of the method in restoring distorted text lines in document images against their original state, we compare it to state-of-the-art algorithms on the widely-used CBDAR 2007 and personal dataset. With SSIM of 0.83 and PSNR of 65.21dB, the proposed approach is found better at dewarping highly distorted English document images. The results on a wide range of distorted document images show that the proposed methodology can be extended to address other document enhancing tasks.