Weichao Chen

Single image deraining aims to recover clean background scenes from raincontaminated images. Existing methods often enlarge the receptive field tomodel long-range spatial dependencies and better capture rain streak patterns. However, many of them rely on Transformer-based self-attention orconventional large-kernel convolutions, whose computational and parametercosts grow rapidly with the receptive field, limiting efficiency and practicality. To address this issue, we propose a Multi-Scale Large Kernel AttentionNetwork for efficient and high-quality single image deraining. Specifically,we design a Large Kernel Attention Block that combines strip convolutionsand dilated convolutions within an attention framework, enabling near-globalcontext aggregation with low computational overhead. To alleviate detaildistortion caused by long-range modeling, we further introduce a Multi-scaleProgressive Detail Recovery module, which progressively fuses cross-scalefeatures and suppresses residual degradation for fine-detail restoration. Inaddition, an Edge Recovery Block is incorporated to enhance horizontal andvertical structural cues, improving edge consistency and overall structuralfidelity. Extensive experiments on both synthetic and real-world benchmarkdatasets demonstrate that MLKANet achieves competitive deraining performance while maintaining favorable efficiency.

Single image deraining aims to restore clean background scenes from rainy images. To address the high computational and parameter costs caused by large receptive field modeling in existing methods, we propose a multi-scale large kernel attention network for efficient and high-quality deraining. Specifically, a large kernel attention block is designed by combining strip convolutions and dilated convolutions within an attention framework, enabling near-global context aggregation with low computational overhead. To alleviate detail distortion introduced by long-range modeling, a multi-scale progressive detail recovery module is further developed to gradually fuse cross-scale features and suppress residual degradations. In addition, an edge-aware reconstruction module is incorporated to enhance horizontal and vertical structural information, improving edge consistency and overall structural integrity. Experimental results show that the proposed method achieves competitive performance on both synthetic and real-world benchmark datasets.