Multi-scale GroupMamba for spectral super-resolution from single RGB image

Spectral super-resolution (SSR) aims to improve the spectral resolution of single RGB image to reconstruct hyperspectral image (HSI). Recent SSR methods are unable to effectively capture the complex correlations between pixels and frequency bands. In this paper, we propose a multi-scale GroupMamba (MSGM) for SSR. Specifically, MSGM is composed of multiple ConMamba blocks, each including a multi-scale spatial-spectral feature extraction (MSF) block, a large-kernel convolution (LKC) block and a spatial-spectral GroupMamba block. In the MSF block, three convolution branches are utilized to extract spatial-spectral information at different scales for capturing multi-level information ranging from details to the global view. In the LKC block, we use large-kernel convolutions in the horizontal and vertical directions respectively to expand the receptive field. Meanwhile, we integrate the spectral-attention block to improve the local feature extraction ability. In GroupMamba block, the spatial GroupMamba (SpaGroup) and spectral GroupMamba (SpecGroup) blocks are used to process the information between adjacent pixels and bands, respectively. The former captures local-to-global spatial associations through sequence reorganization by scanning, while the latter adopts a band grouping strategy to model the cross-band non-linear mapping. We conducted extensive experiments on four benchmark datasets, including CAVE, Harvard, ARAD_1K, and Houston, demonstrating that MSGM outperforms comparative methods. For instance, compared to the suboptimal comparative method, our approach achieves a 5.4% reduction in RMSE and a 1.3% reduction in SAM on the Harvard dataset, respectively.