10th WorldS4 2026

Authors - Xinyi ZHU, Han Wang, Jun WU
Abstract - In rainy-day images, rain patterns exhibit complex morphologies and significant variations in scale, and they tend to overlap with background textures and edge structures, posing significant challenges for rain removal tasks based on a single image. Addressing the shortcomings of existing methods in modeling complex rain patterns, enhancing key regions, and utilizing complementary information across channels, this paper proposes a multi-scale attention and multi-channel fusion image rain removal method tailored for complex rain pattern scenarios. First, we construct a parallel multi-scale feature extraction module that uses standard convolutions and dilated convolutions with varying dilation rates to capture fine-scale local rain patterns, mesoscale rain patterns, and large-scale contextual information, thereby enhancing the network's ability to perceive rain patterns across multiple scales. Second, an adaptive attention optimization module is designed to re-calibrate multi-scale features across both channel and spatial dimensions, enabling the model to focus more on areas with dense rain patterns, edge structures, and effective texture information. Finally, a multi-channel feature interaction and fusion mechanism is introduced. Through channel segmentation, cross-channel interaction, and residual fusion, this mechanism enhances information complementarity between different feature subspaces, thereby improving the structural preservation and visual naturalness of the restored results. Experimental results on the Rain100H, Rain100L, Rain12, and SPA-Data datasets demonstrate that our method achieves superior rain removal performance across multiple test scenarios, exhibiting particularly strong generalization capabilities in real-world rainy conditions.