Real-World Video Super-Resolution with a Degradation-Adaptive Model

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Video super-resolution (VSR) faces challenges with complex, unknown real-world degradations.
• Current VSR methods lack adaptability to varying degradation levels, limiting their practical application.

Purpose of the Study:

• To develop a flexible VSR network capable of handling diverse real-world video degradation levels.
• To improve the accuracy and robustness of VSR for corrupted video sequences.

Main Methods:

• Proposed a degradation-adaptive video super-resolution network (DAVSR) utilizing a bidirectional propagation approach.
• Incorporated a pre-cleaning module for noise and artifact reduction and an unsupervised optical flow estimator for precise inter-frame alignment.
• Adapted network architecture for streamlined propagation and reconstruction.

Main Results:

• DAVSR demonstrated superior performance across diverse degradation types.
• Achieved an average improvement of 0.18 dB over state-of-the-art methods (DBVSR) in PSNR.
• Effectively handled real-world video sequences with varying degradation levels.

Conclusions:

• The proposed DAVSR network offers a flexible and effective solution for real-world VSR problems.
• The degradation-adaptive approach significantly enhances VSR performance on diverse corrupted videos.
• DAVSR shows promise for practical applications requiring high-quality video reconstruction from degraded sources.