Degradation-Aware Transformer for Single Image Deraining

Abstract

The crux of image deraining originates from recognizing various rain patterns. Most existing methods employ low-level spatial or sequential information to reconstruct the rain-free background image. However, due to the lack of sufficiently capturing long-term contextual relation between pixels, these methods often lead to incompletely modeling rain patterns and visible rain residues remained. In this paper, we propose a novel Degradation-Aware Transformer (DAT), which leverages a multi-level contrastive learning to obtain discriminative degradation representations by a degradation-aware model. Based on this, we also design a degradation-aware self-attention mechanism to improve the restoration performance on diverse rain patterns. Benefiting from the developed self-attention mechanism, DAT is able to capture long-term relations between pixels as well as completely modeling the rain degradation patterns. Extensive experiments demonstrate that our DAT is able to achieve the state-of-the-art performance of single image deraining in terms of both qualitatively visual comparison and quantitative comparison with other baseline methods on benchmark datasets.

Keywords

• Image deraining
• degradation representation
• contrastive learning
• generative model