Spatial-Temporal Synergic Prior Driven Unfolding Network for Snapshot Compressive Imaging

In order to develop a fast and accurate algorithm for snapshot compressive sensing (SCI), we combine the merits of two kinds of existing SCI methods: the interpretability of traditional model-based methods and the speed of learning-based ones. Concretely, we build a novel Spatial-Temporal synErgic Prior driven unfolding network for SCI, dubbed STEP-SCI, which is inspired by Half-Quadratic Splitting (HQS) for optimizing the SCI reconstruction model. To better utilize the correlations among frames, we develop a spatial-temporal synergic prior module for solving the proximal mapping problem, which explicitly exploits the temporal correlation and the spatial correlation in turn. All parameters in STEP-SCI are learned end-to-end. Extensive experiments demonstrate that our STEP-SCI has superior performance than existing state-of-the-art model-based and learning-based methods while maintaining fast inference speed.