MHS-Net: A multi-scale heterogeneous synergistic network for single image deraining

Abstract

Single-image rain removal plays a crucial role in improving downstream visual tasks under adverse weather conditions. However, existing methods often fail to effectively balance global–local feature interactions and adaptive feature fusion in complex rain-streak scenarios. To address these challenges, we propose a Multi-scale Heterogeneous Fusion Network (MHF-Net) that integrates three core innovations for enhanced rain removal. The first innovation is the Heterogeneous Synergistic Enhancement (HSE) module, which combines Vision Mamba and convolutional branches to jointly model long-range dependencies and restore fine-grained textures. The second is the Dynamic Perception Adaptive Fusion (DPAF) strategy, which utilizes learnable masks to spatially separate features, reducing fusion artifacts and improving color consistency. Lastly, the Hierarchical Multi-scale Integration Mechanism (HMIM) refines cross-scale features using a pyramid encoder–decoder architecture. On the Rain100L dataset, our approach achieves a notable PSNR improvement of 0.42 dB and elevates SSIM to 0.991, surpassing the state-of-the-art methods. For the more challenging Rain100H dataset, all evaluation metrics show consistent improvements. Visual and residual analyses confirm superior rain removal and detail preservation, while downstream applications, such as semantic segmentation, further demonstrate the practical benefits of the proposed method. Ablation studies validate the contribution of each module in enhancing the overall performance.