Enhancing Single Image De-raining with a Sparse Spatial Transformer

The paper proposes an innovative Transformerbased architecture for rain removal, called the Sparse Spatial Transformer (DRMformer), aimed at effectively enhancing the performance of single image de-raining. DRMformer is a model designed to improve the quality of image reconstruction in single image de-raining tasks. The model introduces two core modules: the Multi-branch Top-k Attention (MTKA) and the Multi-scale Spatial Attention Feed-forward Network (MSAFN). MTKA combines local and global attention mechanisms, using a Top-k selection strategy to effectively filter out interfering information, thereby improving the efficiency of feature aggregation. MSAFN employs multi-scale convolution and spatial attention mechanisms to capture image features at different scales, enhancing the retention of detail in image reconstruction. Experimental results show that DRMformer performs exceptionally well on multiple benchmark datasets, such as Rain200L, Rain200H, DID-Data, and DDN-Data, outperforming current state-of-the-art de-raining methods. Ablation studies further verify the significant contribution of the MTKA and MSAFN modules to the model's performance. DRMformer provides a new solution to the single image de-raining problem, offering significant benefits for applications requiring image processing in rainy conditions and paving the way for future research and application in image processing tasks.