Dual Hyperspectral Mamba for Efficient Spectral Compressive Imaging

TL;DR

This work tackles the CASSI-based hyperspectral image reconstruction problem by introducing Dual Hyperspectral Mamba (DHM), a multi-stage unfolding framework that jointly models global long-range and local textures through dual hyperspectral S4 blocks (GHSB and LHSB) within dual hyperspectral S4 blocks (DHSB). The method learns degradation-pattern parameters $(\eta_t, \rho_t)$ to adaptively scale the linear projection and set the denoiser noise level, enabling efficient and interpretable reconstruction via $T$ iterative stages. Empirical results on simulated KAIST/CAVE data and real measurements show that DHM achieves state-of-the-art PSNR/SSIM with lower model size and FLOPs, and a lightweight DHM-light variant offers strong performance with reduced complexity. The approach advances HSI reconstruction by balancing global context modeling and local texture preservation in a physically informed, learnable unfolding framework.

Abstract

Deep unfolding methods have made impressive progress in restoring 3D hyperspectral images (HSIs) from 2D measurements through convolution neural networks or Transformers in spectral compressive imaging. However, they cannot efficiently capture long-range dependencies using global receptive fields, which significantly limits their performance in HSI reconstruction. Moreover, these methods may suffer from local context neglect if we directly utilize Mamba to unfold a 2D feature map as a 1D sequence for modeling global long-range dependencies. To address these challenges, we propose a novel Dual Hyperspectral Mamba (DHM) to explore both global long-range dependencies and local contexts for efficient HSI reconstruction. After learning informative parameters to estimate degradation patterns of the CASSI system, we use them to scale the linear projection and offer noise level for the denoiser (i.e., our proposed DHM). Specifically, our DHM consists of multiple dual hyperspectral S4 blocks (DHSBs) to restore original HSIs. Particularly, each DHSB contains a global hyperspectral S4 block (GHSB) to model long-range dependencies across the entire high-resolution HSIs using global receptive fields, and a local hyperspectral S4 block (LHSB) to address local context neglect by establishing structured state-space sequence (S4) models within local windows. Experiments verify the benefits of our DHM for HSI reconstruction. The source codes and models will be available at https://github.com/JiahuaDong/DHM.

Figures (6)

• Figure 1: Comparisons of PSNR-FLOPS between our DHM and SOTA models.
• Figure 2: Our unfolding framework with $T$ iterative stages.
• Figure 3: Algorithmic pipeline of our DHM. (a) Architecture of our DHM at the $t$-th iterative stage. (b) Each DHSB is composed of a GHSB, a LHSB, a GFFN and three LN layers. (c) Diagram of the GHSB and LHSB modules. (d) Components of the GFFN. (e) Design of the HSI-SSM.
• Figure 4: Qualitative results on the Scene 7 (S7) of simulation dataset (zoom in for a better view).
• Figure 5: Qualitative comparisons on the Scene 4 (S4) of real dataset (zoom in for a better view).