A comparison of Generative Adversarial Networks for image super-resolution

This article presents a comparison of Generative Adversarial Networks for the image super-resolution problem. This is a relevant problem in several research areas and many real-world applications. The research consists of four steps: selecting successful Generative Adversarial Networks architectures, implementing two promising models, evaluating their image quality results, and analyzing their transfer learning capabilities. The main results indicate that both models are able to compute accurate results, with a reasonable deviation from state-of-the-art results and good transfer capabilities.

I. Introduction

Image super-resolution is a relevant problem in image processing [1]. The main goal of the problem is building or recovering a high resolution (HR) image from a single (or a set of) low-resolution (LR) image(s). Obtaining such super-resolution image is a challenging problem, for which many computational solutions have been proposed [2]. The problem is very important for many research areas and real world applications that rely on specific information on images (e.g., computer vision, medical imaging and assistance, satellite imagery, privacy and security, etc.).