Degradation-Adaptive Denoising: Aligning Diffusion Models With Physics of Video Snapshot Compressive Imaging

Area of Science:

• Computational Imaging
• Deep Learning for Imaging
• Video Reconstruction

Background:

• Video Snapshot Compressive Imaging (SCI) captures high-speed scenes efficiently.
• Current SCI methods use feedforward networks with fixed denoising, limiting physical model alignment and balancing motion/static details.

Purpose of the Study:

• To propose the first diffusion-based reconstruction framework, PCD-Diffusion, for Video SCI.
• To address limitations in existing SCI methods regarding physical model alignment and reconstruction quality.

Main Methods:

• Developed a Physically-Constrained Dynamic Diffusion (PCD-Diffusion) model for Video SCI.
• Introduced a region-adaptive diffusion schedule and spatiotemporal residual estimation.
• Incorporated motion prior-guided diffusion and Gauss-guided residual estimation for dynamic steering.

Main Results:

• PCD-Diffusion explicitly aligns denoising with SCI's residual distribution.
• The framework achieves superior reconstruction fidelity and temporal coherence.
• Demonstrated effectiveness on both simulated and real-world datasets.

Conclusions:

• PCD-Diffusion offers a significant advancement in Video SCI reconstruction.
• The physically-constrained diffusion approach improves motion detail recovery and background smoothing.
• The proposed method provides more accurate and temporally coherent reconstructions compared to existing techniques.