Improved Image Deblurring Using GANs

Abstract

This paper proposes a new generative adversarial network that attains statistically significant improvements over some current image restoration techniques and traditional image restoration algorithms. Most of the existing image restoration techniques depend on prior information about the image. More flexibility can be obtained by using data-driven techniques to restore the image. Data-driven techniques like deep learning (DL) algorithms along with generative adversarial networks (GAN) have been showing great potential in the domain of computer vision (CV). However, it still has a scope of improvement in image denoising, deblurring, and super-resolution as the basic formulation of generative adversarial networks (GANs) cannot be directly used to generate realistic high-resolution images and some structures of the estimated images are usually not preserved well. Images from domains such as medicine and crime highly depend on the reality of the image and cannot entertain false positives and false negatives in their analysis. Such domains are still facing challenges to obtain true better quality images with the existing techniques. This paper also intends to solve some of these problems by giving better outputs with the proposed method. The improvised GAN proposed in this paper uses image quality parameter SSIM to improve restoration.