Region Guided Transformer for Single Image Raindrop Removal

Raindrops can significantly degrade image quality by introducing unwanted reflections and occlusions. Although deep learning-based methods have shown promise in removing these artifacts, they often struggle to completely eliminate raindrop traces and restore the original scene. To address this challenge, we propose a novel Region Guided Transformer Network (RGTN) for single-image raindrop removal. Our RGTN incorporates a unique attention mechanism called Mask-Window Multi-head Self-Attention (MW-MSA), which utilizes a degraded region mask to selectively process the degradations and focus on clean background information. Extensive experiments conducted on benchmark raindrop removal datasets demonstrate the superiority of our RGTN network over existing methods. Our approach achieves more detailed and realistic results. To further validate the generalization ability and robustness of our network, we also evaluate its performance on other related tasks, such as snow removal, using datasets like Snow 100K. The results indicate that our network outperforms the latest methods in this domain as well.