Ev-DeblurVSR: A Tool for Video Deblurring and Super-Resolution Enhancement 156

What is Ev-DeblurVSR?

Ev-DeblurVSR is a technology that enhances super-resolution reconstruction of blurry video by incorporating event data. It comes from the University of Science and Technology of China, the Hefei Comprehensive National Science Center Artificial Intelligence Research Institute, and the National University of Singapore, among other institutions. It can remove video blur while simultaneously improving video resolution. Related results were showcased at the 2025 AAAI conference.

Features of Ev-DeblurVSR

• Video Deblurring: Can tackle blur in videos, whether caused by motion blur or camera shake, making the video clear.
• Video Super-Resolution: Can transform low-resolution blurry videos into high-resolution clear videos, supporting up to 4x resolution enhancement.
• Real-time Processing: Capable of real-time image and video deblurring, suitable for scenarios requiring immediate results.

Technical Highlights of Ev-DeblurVSR

• Event Camera Data Fusion: It utilizes data from event cameras, which capture moments of light intensity change and provide more precise temporal information.
• Reciprocal Feature Deblurring Module (RFD): Uses intra-frame event motion information to remove blur, while incorporating global scene context to enhance the effect.
• Hybrid Deformable Alignment Module (HDA): Combines inter-frame event information and optical flow to improve the accuracy of motion estimation.
• Performance Optimization: In real-world data testing, Ev-DeblurVSR is more accurate (+2.59dB) than the previous best method FMA-Net, and is also 7.28x faster.

Application Scenarios of Ev-DeblurVSR

• Video Surveillance: Can clean up blurry videos caused by camera shake or other factors, making surveillance footage clearer.
• Motion Capture: Supports high-resolution video needs in sports analysis, virtual reality, and other fields.
• Film Post-Production: Can restore or enhance old video footage.

Frequently Asked Questions

Q: Compared with other event-based BVSR methods, what are the main advantages of the Ev-DeblurVSR network?

A: Ev-DeblurVSR more effectively fuses frame and event information and enhances temporal consistency through the Reciprocal Feature Deblurring Module and the Hybrid Deformable Alignment Module. On the GoPro dataset, its PSNR reaches 32.51, higher than other event-based methods such as EFNet + EGVSR's 23.53; on real data, compared to FMA-Net, it improves accuracy by 2.59dB and is 7.28x faster.

Q: How does the Reciprocal Feature Deblurring Module specifically implement frame feature deblurring and event feature enhancement?

A: The Reciprocal Feature Deblurring Module uses intra-frame event motion information to deblur frame features, while leveraging the global scene context of frames to enhance event features, achieving effective fusion and mutual enhancement of both types of information. The specific implementation involves feature extraction and interactive processing of frames and events.

Q: In the Hybrid Deformable Alignment Module, how do the complementary motion information from inter-frame events and optical flow work together to improve motion estimation?

A: The Hybrid Deformable Alignment Module fully leverages the complementary motion information from inter-frame events and optical flow. By fusing and complementarily calculating both sources of motion information, it more accurately estimates object motion during the deformable alignment process, thereby enhancing temporal consistency.

Project Page:https://dachunkai.github.io/ev-deblurvsr.github.io/

Paper:https://arxiv.org/pdf/2504.13042

Github:https://github.com/DachunKai/Ev-DeblurVSR