Mask-Informed Spatial-Spectral Attention Unfolding Network for Snapshot Compressive Reconstruction

Snapshot compressive imaging (SCI) facilitates the computational recovery of high-dimensional spatio-temporal data through single-shot 2D coded measurements. However, existing SCI reconstruction paradigms usually exhibit inherent limitations in comprehensively exploiting the structural information embedded in optical modulation masks while effectively modeling complex spatial-spectral interdependencies, resulting in suboptimal reconstruction accuracy. To address the above challenges, a novel Mask-Informed Spatial-Spectral Attention Unfolding Network (MSS-Net) is proposed, which is able to systematically explore optical mask priors and model spatial-spectral correlations within a deep unfolding architecture. Specifically, for mask prior exploration, the mask is comprehensively utilized throughout the entire unrolled iterative architecture: a) In the gradient descent process, the mask is employed to jointly generate dynamic gradient and step-size maps, enabling mask-guided fine-grained adaptive gradient updating. b) In the proximal mapping process, the mask provides rich contextual guidance for feature extraction at different scale spaces, facilitating effective mask information fusion and efficient learning of deep representations. On the other hand, for spatial-spectral correlation modeling, an adaptive feature fusion mechanism with a hierarchical encoder-decoder architecture is proposed, which is able to jointly capture fine-grained spatial details and long-range spectral dependencies, thus ensuring high structural fidelity and robust spectral coherence in the reconstructed data. Extensive experiments demonstrate that the proposed MSS-Net outperforms the existing SCI reconstruction approaches on both simulated and real HSI datasets.