Frequency domain task-adaptive network for restoring images with combined degradations

Abstract

Image restoration seeks to obtain a high-quality image by eliminating degradations. While existing methods have shown remarkable performance in image restoration, the majority are designed for single degradation. However, in the real world, environmental factors are complex and variable, leading to images with combined degradation factors—like the simultaneous presence of rain, noise, and haze in a single image. This complexity poses a challenge for existing methods to be effectively applied in real-world scenes. In this paper, we propose a frequency domain-based network for adaptively restoring images with various combinations of degradation factors. Specifically, we design a frequency domain-based gate block (FDGB) to selectively determine which low and high-frequency information should be preserved, choosing the most informative components for recovery. Additionally, we develop a task adaptive block (TAB) composed of FDGBs and frequency domain-based re-weight blocks (FDRBs) to adaptively restore various combined degraded images. FDRB ensures that the TAB can explore various combinations of FDGBs by utilizing a gating mechanism to re-weight the output features of FDGB based on the input signals. Finally, we introduce a Fast Fourier Block (FFB) to enrich the feature set and provide collaborative refinement for the FDGB. To facilitate the training of our proposed method, we create a dataset with various combinations of degradation factors. The resulting tightly interlinked architecture, named as FDTANet, extensive experiments demonstrate that our approach excels not only in restoring images afflicted with combined degradations but also demonstrates competitive performance when compared to state-of-the-art models for single-degradation restoration. The code and the pre-trained models are released at https://github.com/Tombs98/FDTANet/.