A Systematic Review on Generative Adversarial Network (GAN): Challenges and Future Directions

Abstract

Generative adversarial network, in short GAN, is a new convolution neural network (CNN) based framework with the great potential to determine high dimensional data from its feedback. It is a generative model built using two CNN blocks named generator and discriminator. GAN is a recent and trending innovation in CNN with evident progress in applications like computer vision, cyber security, medical and many more. This paper presents a complete overview of GAN with its structure, variants, application and current existing work. Our primary focus is to review the growth of GAN in the computer vision domain, specifically on image enhancement techniques. In this paper, the review is carried out in a funnel approach, starting with a broad view of GAN in all domains and then narrowing down to GAN in computer vision and, finally, GAN in image enhancement. Since GAN has cleverly acquired its position in various disciplines, we are showing a comparative analysis of GAN v/s ML v/s MATLAB computer vision methods concerning image enhancement techniques in existing work. The primary objective of the paper is to showcase the systematic literature survey and execute a comparative analysis of GAN with various existing research works in different domains and understand how GAN is a better approach compared to existing models using PRISMA guidelines. In this paper, we have also studied the current GAN model for image enhancement techniques and compared it with other methods concerning PSNR and SSIM.