Meta-learning based blind image super-resolution approach to different degradations.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Supervised single image super-resolution (SISR) methods require paired training data, limiting real-world application.
• Existing methods struggle with diverse and unknown image degradations.

Purpose of the Study:

• To develop an unsupervised blind SISR method capable of handling various real-world image degradations.
• To enable high-resolution (HR) image restoration from low-resolution (LR) inputs with unknown degradations.

Main Methods:

• Proposed Different Degradations Blind Super-Resolution (DDSR) method using Gaussian modeling for blur degradation.
• Employed a meta-learning framework with random kernel learning for effective blur degradation optimization.
• Utilized alternative optimization between blur degradation and image restoration to handle multiple degradations.

Main Results:

• DDSR demonstrates superior performance compared to state-of-the-art methods on public datasets.
• The method shows improved application flexibility, convenience, and generalization ability across multiple degradations.
• Achieved comparable memory load and time consumption to existing methods.

Conclusions:

• DDSR offers a robust and flexible unsupervised solution for blind SISR in real-world scenarios.
• The plug-and-play implementation allows restoration of HR images from inputs with partial coverage, noise, and darkening.
• The meta-learning approach effectively addresses diverse image degradation challenges.