RASP-Net: Residual Block with High Similarity Pass Attention for Image Desnowing

In snowy seasons or high-altitude regions with frequent snowfall, images are often heavily obscured by snowflakes, severely degrading their visual quality. While existing image desnowing methods typically focus on extracting abundant features to detect and remove multi-scale snowflakes, they often lack effective mechanisms for feature selection and discrimination. To address this limitation, we propose RASP-Net, a Residual Block and High Similarity Pass Attention based Desnowing Network, which effectively reveals image details obscured by snowflakes. Specifically, RASP-Net integrates residual learning with a spatial attention mechanism to dynamically enhance useful features while suppressing irrelevant snowflake noise. It further incorporates a high-similarity attention mechanism combined with soft-thresholding to adaptively identify snowflake features. Finally, a recurrent refinement module is introduced to progressively enhance feature extraction and snowflake removal, thereby improving both the desnowing performance and the robustness of the model. Extensive experiments demonstrate that RASP-Net consistently outperforms state-of-the-art methods on both synthetic and real-world datasets.