Deep Equilibrium Models for Snapshot Compressive Imaging

The ability of snapshot compressive imaging (SCI) systems to efficiently capture high-dimensional (HD) data has led an inverse problem, which consists recovering the HD signal from compressed and noisy measurement. While reconstruction algorithms grow fast solve it with recent advances deep learning, fundamental issue accurate stable recovery remains. To this end, we propose equilibrium models (DEQ) for video SCI, fusing data-driven regularization convergence in a theoretically sound manner. Each model implicitly learns nonexpansive operator analytically computes fixed point, thus enabling unlimited iterative steps infinite network depth only constant memory requirement training testing. Specifically, demonstrate how DEQ can be applied two existing SCI reconstruction: recurrent neural networks (RNN) Plug-and-Play (PnP) algorithms. On variety datasets real data, both quantitative qualitative evaluations our results effectiveness stability proposed method. code are available at: https://github.com/IndigoPurple/DEQSCI.