A NEW BACKBONE FOR HYPERSPECTRAL IMAGE RECONSTRUCTION

As the inverse process of snapshot compressive imaging, the hyperspectral image (HSI) reconstruction takes the 2D measurement as input and posteriorly retrieves the captured 3D spatial-spectral signal. Built upon several assumptions, numerous sophisticated neural networks have come to the fore in this task. Despite their prosperity under experimental settings, it's still extremely challenging for existing networks to achieve high-fidelity reconstructive quality while maximizing the reconstructive efficiency (computational efficiency and power occupation), which prohibits their further deployment in practical applications. In this paper, we firstly conduct a retrospective analysis on aforementioned assumptions, through which we indicate the imminent aspiration for an authentically practical-oriented network in reconstructive community. By analysing the effectiveness and limitations of the widely-used reconstructive backbone U-Net, we propose a Simple Reconstruction Network, namely SRN, just based on some popular techniques, e.g., scale/spectral-invariant learning and identity connection. It turns out, under current conditions, such a pragmatic solution outperforms existing reconstructive methods by an obvious margin and maximize the reconstructive efficiency concretely. We hope the proposed SRN can further contribute to the cutting-edge reconstructive methods as a promising backbone, and also benefit the realistic tasks, i.e., real-time/high-resolution HSI reconstruction, solely as a baseline.