Critical Review on Deep Learning and Smart Technologies for Image Super-Resolution

Image super-resolution is an extremely useful way to improve the quality of an image. It is miracle that making use of the current signal processing and deep learning technologies, the image can look much appealing after the super-solution. This review paper is to highlight important techniques, especially to point out recent key contributions to make superior success of super-resolution of the recent years, especially on face super-resolution. We will start with a very brief and quick review on using conventional signal processing and classic learning approaches for super-resolution, and then concentrate on giving the advantages of deep learning, in particular, the recent powerful concepts on using latent vector and facial priors to achieve superior performance. Further topics of discussion include generative adversarial network (GAN), StyleGAN, latent space, facial priors and diffusion models. Our concentration is on the reasons for the success of these techniques. Attractive demonstrations on a few state-of-the-art models, including some of our work, are provided

I. Introduction

Nowadays, a low resolution image gives us the impression of low-tech, old fashion, awkward style and uncertainty. This is particularly true if the image is used for advertisement, surveillance, medical diagnosis or object recognition. The low quality face image might devalue a person's beauty and faith. In order to relieve the problem we may turn the image into higher resolution, with super-resolution technology. Image super-resolution (SR) usually refers to an increase of the resolution of a single low-resolution (LR) image by up-sampling, deblurring and denoising, while the resultant high-resolution (HR) image should preserve the characteristics of the natural image, such as sharp edges and rich texture.