Augmented Deep Unrolling Networks for Snapshot Compressive Hyperspectral Imaging

Xinran Qin, Yuhui Quan, Hui Ji

Keywords: Hyperspectral Imaging, Snapshot Compressive Imaging, Image Reconstruction, Deep Unrolling Networks

Abstract: Snapshot compressive hyperspectral imaging requires reconstructing a hyperspectral image from its snapshot measurement. This paper proposes an augmented deep unrolling neural network for solving such a challenging reconstruction problem. The proposed network is based on the unrolling of a proximal gradient descent algorithm with two innovative modules for gradient update and proximal mapping. The gradient update is modeled by a memory-assistant descent module motivated by the momentum-based acceleration heuristics. The proximal mapping is modeled by a sub-network with a cross-stage self-attention which effectively exploits inherent self-similarities of a hyperspectral image along the spectral axis, as well as enhancing the feature flow through the network. Moreover, a spectral geometry consistency loss is proposed to encourage the model to concentrate more on the geometric layer of spectral curves for better reconstruction. Extensive experiments on several datasets showed the performance advantage of our approach over the latest methods.

TL;DR: An augmented deep unrolling networks for snapshot compressive hyperspectral Imaging