A Collaborative Network of Mamba and CNN for Lightweight Image Super-Resolution

Single image super-resolution (SR) can recover high-resolution images from the corresponding low-resolution counterparts, which meets the application demands of consumer electronics, such as improving the visual experience of smart televisions (TVs) and virtual reality (VR) devices. Although deep learning-based SR methods have recently gained promising success, most existing approaches typically stack numerous network layers, which significantly increases the model complexity and hinders the deployment on electronics with limited computational capability. To tackle this problem, we propose a collaborative network of Mamba and CNN (CNMC) for lightweight image super-resolution. CNMC is mainly composed of multiple collaborative units of Mamba and CNN (CUMCs), which leverage the complementary advantages of Mamba and CNN to extract deep features beneficial for reconstruction. Specifically, CUMC introduces Mamba which enjoys the global receptive field and linear complexity, to perform long-range dependency modeling. Additionally, it employs CNN with the significant inductive bias to facilitate the local information interaction and compensation. The collaboration between Mamba and CNN effectively exploits both non-local and local priors, ensuring a comprehensive enhancement of deep features. Furthermore, a multi-scale spatial refinement attention (MSSRA) is developed in CUMC, to modulate channel-wise weights of features using a spatial fine-grained manner at different scales and then aggregate these cross-scale features, thereby distilling important information and restoring more precise details. Extensive quantitative and qualitative experiments demonstrate the superiority of our CNMC over other state-of-the-art lightweight SR methods. Most importantly, compared to the recent Transformer-based methods NGswin and HSSRNet, our CNMC achieves PSNR improvements of 0.21 dB and 0.32 dB for x2 SR on Urban 100 dataset. Code available at https://github.com/HITXinWang/CNMC.
© 2025 IEEE