Event-Guided Online Video Super-Resolution

Area of Science:

• Computer Vision
• Image Processing

Background:

• Frame-only video super-resolution (VSR) struggles with motion blur, rapid dynamics, and poor illumination.
• Existing event-guided VSR methods often prioritize reconstruction quality over real-time performance and efficiency.

Purpose of the Study:

• To develop a lightweight and efficient event-guided VSR framework (E2VSR) for real-time applications.
• To enable low-latency event-guided VSR in causal settings using only current and past observations.

Main Methods:

• Proposed an event-confidence adaptive propagation strategy with Event-induced Feature Modulation (EvFM) for cross-modal integration.
• Introduced Event-Confidence Feature Fusion (EvCFF) to leverage events as motion cues for adaptive inter-frame aggregation.
• Implemented Implicit Event Reconstruction (IER) to enrich feature representations during training without inference cost.

Main Results:

• E2VSR achieves superior quantitative and qualitative performance in video super-resolution.
• The framework maintains a low parameter count and computational cost, suitable for real-time deployment.
• Demonstrated improved motion-aware temporal aggregation in challenging dynamic conditions.

Conclusions:

• E2VSR offers an efficient solution for real-time event-guided video super-resolution.
• The proposed methods effectively integrate event and frame data for enhanced VSR.
• E2VSR addresses the trade-off between reconstruction quality and computational efficiency in VSR.