Parameter Free Vision Mamba for Lightweight Image Super-Resolution

Efficient Image Super-Resolution refers to a class of computational methods that enhance image resolution from low-quality inputs to high-definition outputs with optimized resource utilization. Mamba represents an innovative advancement in sequence modeling, leveraging state space models (SSMs) to achieve superior performance in both NLP and time series analysis applications. Its core strength lies in its ability for long-range dependency modeling while maintaining low computational complexity. Despite the remarkable success of Mamba in the field of natural language processing, there have been relatively few studies and applications of Mamba-based methods in the field of efficient image super-resolution in recent years. In this paper, we propose a lightweight image super-resolution network called Parameter Free Vision Mamba (PFVM). PFVM architecture is structured into three key components: shallow feature extraction, deep feature extraction, and a reconstruction module. To enhance residual learning, the Residual State Space Block (RSSB) incorporates 3-residual branches and integrates the Visual State Space Module (VSSM). This design leverages Mamba's long-range dependency modeling to capture richer hierarchical features. In addition, we employ parameter free attention to further enhance the performance of RSSB. Experiments show that our proposed PFVM achieves relatively low complexity while ensuring a certain level of performance. For example, in the Mamba-based methods, PFVM outperforms DVMSR with slightly higher FLOPs and parameters than DVMSR, and while lower than MambaIR in terms of PSNR and SSIM, FLOPs and parameters are only 46% and 60% of those of MambaIR.