AirNet: An All-in-One Image Restoration Network for Multiple Degradations

To tackle the aforementioned problem, we propose an all-in-one image restoration network (AirNet), which consists of two modules. To be specific, Contrastive-Based Degraded Encoder (CBDE) is designed to learn the degradation representation by leveraging the consistency of the images with the same degradation and the inconsistency existing into different degradations. Under the guidance of the degradation representations learned by CBDE, Degradation-Guided Restoration Network (DGRN) aims to restore the images with various degradations. Thanks to the corporation of CBDE and DGRN, AirNet enjoys two highly expected merits, i.e., i) it provides an all-in-one solution to recover the images with different corruption types and ratios; ii) it is free from the prior of the corruption type and ratio. Notably, the referred all-in-one solution is different from existing so-called unified image restoration methods [3,8,23] in given aspects. On the one hand, the methods [3, 8, 23] have to specify the corruption type and ratio, whereas our method does not. On the other hand, they usually treat multiple degradations as a multi-task learning problem with multiple input and output heads, where each input and output head corresponds to a predetermined corruption with a given corruption ratio. In contrast, AirNet is a single pass network which does not differentiate different corruption types and ratios, thus enjoying better flexibility and higher economy.

To summarize, the contribution and novelty of this study are as below:

• As far as we know, AirNet could be one of the first methods to recover images from multiple corruptions in an all-in-one fashion. As our method does not require any degradation information for restoration in advance, it might be closer to the real world scenario.
• AirNet works in a dual manner, which contrastively learns the degradation representation from the observed images and then uses the learned degradation representation to restore the clean image. It should be pointed out that the success of contrastive learning heavily relies on the construction of positive and negative pairs. In this paper, we show a novel method that is effective to capture the inherited characteristics of multi-degradations.
• Without loss of generalizability, we conduct extensive experiments to verify the effectiveness of AirNet in denoising, deraining and dehazing, comparing with 17 baselines.