Multiscale Interactive Spectral Attention Networks for Spectral Super-Resolution

Spectral super-resolution (SSR) focuses on generating a hyperspectral image (HSI) from a single RGB image. While deep learning has shown great promise in SSR, existing methods often overlook spatial features and fail to effectively capture inter-band relationships, resulting in suboptimal reconstructions. To address these challenges, we propose MISANet, a Multiscale Interactive Spectral Attention Network. Inspired by inter-frame utilization in video super-resolution, MISANet employs two U-Net-like branches and integrates advanced attention mechanisms to extract spectral information. Extensive experiments on the TG1HRSSC dataset show that MISANet surpasses state-of-the-art SSR methods in both qualitative and quantitative evaluations.