SSRMamba: Efficient Visual State Space Model for Spectral Super-Resolution

Abstract

Spectral super-resolution, which reconstructs hyperspectral images (HSI) from a single RGB image, has garnered increasing attention. Due to the limitations of CNN structures in spectral modeling and the high computational cost of Transformer structures, existing deep learning (DL)-based methods struggle to balance spectral reconstruction quality and computational efficiency. Recently, Mamba methods base on state-space models (SSM) show great potential in modeling long-range dependencies with linear complexity. Therefore, we introduce the Mamba model into spectral super-resolution (SSR) task. Specifically, we propose a three-stage SSR network base on Mamba, called SSRMamba. We design SpaMamba, SSMamba, and SpeMamba modules for shallow spatial information extraction, mixed information encoding, and spectral information reconstruction, respectively. Extensive experimental results demonstrate that SSRMamba not only surpasses existing methods in terms of quantification and quality, achieving state-of-the-art (SOTA) performance, but also significantly reduces model size and computational cost. The source code of SSRMamba is available at: https://github.com/Baisonm-Li/SSRMamba.