Benchmark of deep learning models for single image super-resolution (SISR)

In this paper we present a study of deep learning models for single image super-resolution (SISR), through the some latest methods used in neural networks for super-resolution, exploring many methods used and proposed. Moreover, this paper presents a global and complete technical benchmark of state-of-the-art machine learning algorithms based on reference metrics (PSNR and SSIM) in the field of visualization and perception. This study involved 53 different neural networks tested on 7 datasets established as reference in the vision domain (Set5, Set14, BSD100, Urban100, DIV2K, Manga109, DIV8K). This work allows us to have a reference to demonstrate the performances and the limits of these algorithms as well as to orient future research in the field of super resolution images in order to develop efficient algorithms. The benchmark covered many neural network architectures (GAN, RNN and Residual Networks), using different techniques and distinct technologies.

I. Introduction

Benchmarking is a tool for analysis and decision making that allowed to measure the performance of existing and possible technical solutions. This methodological approach allows us in this article to orient the selection of existing super-resolution models, or to improve the architectures and techniques for the design and implementation of deep learning models to better achieve the objective of super-resolution of images. In this sense, we incorporate and adapt the best practices of the super-resolution domain based on deep learning methods, not by imitating existing architectures, but by exploring them and studying their technical and architectural constraints.