BVSR-EvD: Blurry Video Space-Time Super-Resolution With Events via Diffusion Models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Video restoration from low-resolution, low-frame-rate, and blurry sources is difficult due to limited data.
• Existing methods struggle with insufficient data priors for effective restoration.

Purpose of the Study:

• To propose BVSR-EvD, a novel approach for blurry video space-time super-resolution.
• To leverage event cameras and diffusion models for enhanced video restoration.

Main Methods:

• BVSR-EvD utilizes event-video dual modalities to extract three data priors: motion (events), content (videos), and physical (integration).
• The Trident Diffusion Model (Trident-DM) decomposes denoising into trident decoupling and adaptive self-composition stages.
• Meta-networks extract priors, and learned weight maps dynamically integrate them.

Main Results:

• BVSR-EvD achieves up to \times 8 spatial super-resolution and \times 64 temporal super-resolution.
• The method demonstrates superior performance compared to existing techniques on public video datasets.
• The integrated priors enhance temporal stability, content preservation, and detail.

Conclusions:

• BVSR-EvD effectively addresses the challenges of blurry video restoration.
• The proposed method offers significant improvements in both spatial and temporal super-resolution.
• This work highlights the potential of combining event cameras and diffusion models for advanced video processing.