Feature decoupling and reorganization network for single image deraining

Single image deraining has become an important preprocessing task in the multimedia area, improving the performance of subsequent high-level computer vision tasks in rainy weather significantly. Previous efforts to restore rain-damaged images have mostly concentrated on feature extraction and contextual information mining. The fact that rain streaks and the background are frequently intermingled during the deraining process is ignored by the majority of these methods, which leads to redundant features and somewhat restricts the generalizability of deraining networks in real-world rainy situations. To address this issue, we present a novel multi-scale single image deraining network called Feature Decoupling and Reorganization Network (FDRNet). FDRNet introduces Dilated Pyramid Split Attention Module (DPSAM) to decouple input features and reorganize extracted features. Additionally, we propose a Dual Prior Fusion Module (DPFM) to extract and fuse prior information from images, which is demonstrated to support the preservation of structural information and the restoration of details. Extensive experiments on both synthetic and real-world datasets show that our proposed method achieves state-of-the-art results.