Answers are grounded in this paper and cite their sources.
Video snapshot compressive imaging (SCI) enables high-speed video acquisition by encoding multiple temporal frames into a single coded measurement. However, recovering the original video sequence from compressed observations remains a challenging ill-posed inverse problem, requiring effective integration of physical measurement constraints and spatio-temporal prior learning. In this work, we propose LDUnet, an efficient lightweight deep unfolding framework for video SCI reconstruction. LDUnet incorporates the SCI measurement model into the unfolding process through a measurement-normalized fidelity update and employs a compact spatio-temporal prior with parameter sharing across unfolding stages. Extensive experiments on simulated grayscale and color SCI datasets demonstrate that LDUnet achieves competitive reconstruction performance with substantially reduced model complexity. Additional ablation experiments verify the contribution of measurement normalization and show that stage-wise parameter sharing substantially reduces model size with only a marginal change in reconstruction accuracy. Evaluations under unseen masks, different measurement noise levels, and real SCI measurements with different compression factors further characterize the reconstruction behavior of LDUnet under varying sensing conditions. These results indicate that LDUnet provides an efficient measurement-consistent unfolding solution for video SCI reconstruction, achieving a favorable balance between reconstruction quality, model complexity, and computational efficiency.
No takes yet. Share an insight, caveat, or question.
Hu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: