Deep image enhancement

Abstract

Deep-learning based methods have brought a huge improvement in the field of image restoration and enhancement. Recent methods explore generative priors from pre-trained generator such as StyleGAN for the task of restoration. In this work, I follow this direction and delve deeper to gain more insights. I first conduct experiments and analysis on a relatively mature task – image denoising. My experiments demonstrate that the generative priors encapsulated in a generative network (StyleGAN) is able to improve the performance in not only super-resolution but also denoising. Furthermore, I analyze the sensitivity of such networks toward the changes of the input image. I find that even a subtle change in the input could lead to substantial changes in the output. Motivated by my findings, I shift the focus to the task of real-world face image restoration, and I devise a simple yet effective image manipulation method that could largely improve the performance of the outputs of a pre-trained model.