Abstract: In photographs taken in snowy environments during winter or at high altitudes, a large number of snowflakes obscure the images, severely degrading their quality. Current image desnowing algorithms mainly focus on extracting massive features to detect and remove multi-scale snowflakes, but they lack an effective mechanism for feature discrimination and selection. To address this issue and extract snowflake features with stronger relevance, we propose an image desnowing network called RASP-Net, which is based on residual blocks and high-similarity propagation attention. First, residual blocks are combined with a spatial attention mechanism to dynamically enhance useful features while suppressing irrelevant snowflake features. Then, high-similarity propagation attention and soft-thresholding techniques are employed to adaptively identify snowflake features. Finally, a recurrent structure is introduced to refine feature extraction and snowflake removal, thereby improving the desnowing performance and robustness of the model. Extensive experiments on the Snow100K, SRRS, and CSD datasets demonstrate that RASP-Net achieves superior desnowing results on both synthetic and real-world datasets compared to state-of-the-art methods. In particular, on the Snow100K dataset, RASP-Net achieves a peak signal-to-noise ratio (PSNR) of 35.51 dB, and can accurately reconstruct image details lost due to snowflake occlusion.