Blind Super-Resolution via Meta-Learning and Markov Chain Monte Carlo Simulation

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Blind single image super-resolution (SISR) traditionally requires predefined kernel priors.
• Existing learning-based methods often necessitate handcrafted or learned kernel priors, limiting adaptability.

Purpose of the Study:

• To develop a novel blind SISR approach that learns kernel priors autonomously.
• To introduce a plug-and-play, unsupervised inference solution for SISR.

Main Methods:

• Utilizing meta-learning and Markov Chain Monte Carlo (MCMC) simulations with random Gaussian distributions to learn kernel priors.
• Employing a lightweight network as a kernel generator, optimized with network-level Langevin dynamics to avoid local optima.
• Implementing a meta-learning-based alternating optimization for kernel generator and image restorer, enhancing convergence.

Main Results:

• The proposed method successfully learns effective kernel priors from organized randomness.
• Network-level Langevin dynamics prevent suboptimal kernel estimations.
• Meta-learning based optimization yields improved convergence compared to traditional methods.
• The approach demonstrates superior performance and generalization on synthetic and real-world datasets.

Conclusions:

• The developed approach offers a learning-based, plug-and-play solution for unsupervised blind SISR.
• It effectively learns kernel priors, overcoming limitations of traditional methods.
• The technique shows significant potential for advancing image super-resolution tasks.