Self-Supervised Deep Blind Video Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Supervised deep learning for video super-resolution (SR) relies on simplified degradation models (e.g., bicubic kernel), which fail in real-world scenarios.
• Acquiring paired low-resolution (LR) and high-resolution (HR) video data for training is challenging in practical applications.
• Existing methods struggle with complex, unknown degradation processes in real-world videos.

Purpose of the Study:

• To develop a self-supervised learning method for blind video SR that addresses limitations of supervised approaches.
• To enable simultaneous estimation of blur kernels and HR videos directly from LR videos.
• To improve video SR performance on complex, real-world degradation scenarios.

Main Methods:

• Proposing a self-supervised learning framework for blind video SR.
• Developing a method to generate auxiliary paired data from LR videos for network constraint.
• Integrating an optical flow estimation module to leverage temporal information from adjacent frames.
• Simultaneously estimating blur kernels and restoring latent HR videos.

Main Results:

• The proposed method effectively handles complex and unknown video degradations.
• Generated auxiliary data provides robust constraints for blur kernel estimation and HR video restoration.
• Incorporation of optical flow enhances the exploitation of temporal information.
• Experimental results demonstrate superior performance compared to state-of-the-art methods on benchmarks and real-world videos.

Conclusions:

• The self-supervised approach overcomes the need for paired data and idealized degradation models in video SR.
• The method offers a practical solution for blind video SR in real-world applications.
• The technique shows significant potential for enhancing video quality in diverse scenarios.