Towards Universal Deep Learning Models for Image Restoration

Abstract

Recent years have witnessed the remarkable successes of deep learning methods in the field of image restoration. However, despite the similarities across different image restoration tasks, researchers often adopt a problem-specific approach. Most deep learning based image restoration algorithms are tailored to a specific type of degradation, performing poorly when being applied to degradations that are deviated from those of the training dataset. This lack of universality limits the adaptability and robustness of these algorithms in real-world scenarios. The approach of training and storing multiple models for various degradation types wastes resources and reduces efficiency, and it still tends to struggle with unseen and complex degradation sources. In this thesis, we depart from current problem-specific methodologies for image restoration and strive to improve the universality and robustness of the existing methods. We propose three novel methods to achieve the above goal; they are a new inference method, a new network model, and a new training method, respectively.