Dynamic frequency window transformer for single image deraining

Abstract

Outdoor visual systems often encounter rain streaks leading to the degradation of images that affects later high-level tasks. In contrast to methods that currently perform feature extraction in the spatial domain, this paper proposes a Dynamic Frequency Window Transformer Network(DFWT-Net) by integrating the frequency and spatial domain information of images. Firstly, the Space Frequency Module (SFM) is utilized to extract local features and preliminary frequency characteristics of the image, including amplitude and phase. Subsequently, the residuals of the rain streaks are obtained by a hierarchical encoder–decoder network consisting of dynamic frequency window transformers (DFWTs). The Dynamic Frequency Window Transformer (DFWT) employs learnable masks for dynamic frequency filtering before self-attention computation to emphasize the frequency of rain streaks. To obtain rain-free images that resemble real scenes, this paper propose the Cosine Phase Loss (CPL) function to measure the phase similarity between the recovered image and the ground truth image. The experimental results thoroughly validate the effectiveness and robustness of the proposed network.