Abstract
Hyperspectral imaging is an essential imaging modality for a wide range ofapplications, especially in remote sensing, agriculture, and medicine. Inspiredby existing hyperspectral cameras that are either slow, expensive, or bulky,reconstructing hyperspectral images (HSIs) from a low-budget snapshotmeasurement has drawn wide attention. By mapping a truncated numericaloptimization algorithm into a network with a fixed number of phases, recentdeep unfolding networks (DUNs) for spectral snapshot compressive sensing (SCI)have achieved remarkable success. However, DUNs are far from reaching the scopeof industrial applications limited by the lack of cross-phase featureinteraction and adaptive parameter adjustment. In this paper, we propose anovel Hyperspectral Explicable Reconstruction and Optimal Sampling deep Networkfor SCI, dubbed HerosNet, which includes several phases under theISTA-unfolding framework. Each phase can flexibly simulate the sensing matrixand contextually adjust the step size in the gradient descent step, andhierarchically fuse and interact the hidden states of previous phases toeffectively recover current HSI frames in the proximal mapping step.Simultaneously, a hardware-friendly optimal binary mask is learned end-to-endto further improve the reconstruction performance. Finally, our HerosNet isvalidated to outperform the state-of-the-art methods on both simulation andreal datasets by large margins. The source code is available athttps://github.com/jianzhangcs/HerosNet.