This paper introduces an all-in-one image restoration network capable of handling multiple image degradations. The proposed system uses a neural degradation representation (NDR) to understand and decompose various types of degradations. A degradation query module and a degradation injection module are developed to recognize and utilize the specific degradation based on NDR, enabling restoration ability for multiple degradations. The paper also proposes a bidirectional optimization strategy to drive NDR to learn the degradation representation by optimizing the degradation and restoration processes alternately. The method demonstrates effectiveness and generalization capability on representative types of degradations such as noise, haze, rain, and downsampling.
Publication date: 20 Oct 2023
Project Page: ?
Paper: https://arxiv.org/pdf/2310.12848
