High-resolution image deraining via dual-branch features interaction and fusion

Existing image deraining methods primarily focus on low-resolution images and exhibit inherent limitations when applied to high-resolution images. First, performing feature extraction directly on high-resolution images leads to a significant increase in computational cost and memory consumption. Second, the inherent complexity of high-resolution rainy images, where rain streaks often intermingle with or severely obscure the background, imposes substantial demands on models' capacity to capture both local features and global contextual dependencies. Consequently, many existing models tend to over-smooth high-frequency details during rain removal, resulting in image blurring and the loss of critical structural information. To address these challenges, we propose an efficient Dual-branch Deraining Network (DDNet), which integrates frequency-domain feature interaction and dynamic feature fusion to enhance the representations capacity of its CNN and Transformer branches. Specifically, we propose an Adaptive Frequency Reciprocal Module (AFRM) that adaptively generates frequency spectral filters to enable mutual reinforcement and interaction between the two branches. We further propose a Discrete Cosine Transform (DCT) based Multi-Spectral Fusion Module (MSFM) to dynamically fuse the decoded features from both branches, thereby promoting effective aggregation of multi-spectral information. Moreover, to address the potential feature disparities and uncertain knowledge gaps between the CNN and Transformer architectures, we propose a Dual Differential Cross Attention Module (DDCAM) to mitigate the adverse effects stemming from feature discrepancies. Extensive experiments conducted on both real-world and synthetic high-resolution rainy datasets demonstrate the effectiveness and efficiency of our model. Our code and models are available at: https://github.com/YT3DVision/HRDerain.