Geometry-Aware Deep Video Deblurring via Recurrent Feature Refinement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Video blurring is common due to camera shake and motion.
• Traditional methods struggle with complex blurring, while deep learning lacks geometric priors.
• Existing deep methods fail to address extreme blurring from significant motion or depth variations.

Purpose of the Study:

• To develop a geometry-aware deep video deblurring framework.
• To integrate optimization-based and deep-learning approaches for improved deblurring.
• To enhance the handling of extreme blurring scenarios.

Main Methods:

• A recurrent feature refinement module is proposed.
• This module fuses deep video features with geometrical information.
• It iteratively refines both features and geometry for precise latent frame restoration.

Main Results:

• The proposed framework was tested on eight baseline networks.
• Experimental results demonstrate superior performance compared to baselines.
• The method achieves state-of-the-art results on four benchmark datasets.

Conclusions:

• The geometry-aware deep video deblurring method effectively restores latent frames.
• The recurrent refinement and fusion of geometric information are key to its success.
• This approach significantly advances video deblurring capabilities, especially for challenging cases.