Bayesian Image Super-Resolution With Deep Modeling of Image Statistics.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep learning methods have largely overlooked the modeling of image prior statistics for super-resolution.
• Natural image statistics, combining smoothness and sparsity, are crucial for effective image restoration.

Purpose of the Study:

• To propose a Bayesian image restoration framework incorporating smoothness and sparsity priors.
• To develop a variational Bayesian approach for inferring image priors.
• To implement and evaluate an unsupervised training strategy for single image super-resolution (SISR).

Main Methods:

• Modeled an ideal image as a sum of smoothness and sparsity components.
• Incorporated realistic image degradations like blurring, downscaling, and noise.
• Developed a variational Bayesian inference method to estimate posterior distributions.
• Implemented the framework for SISR using deep neural networks with an unsupervised training strategy.

Main Results:

• Demonstrated superior model generalizability across varying noise levels and degradation kernels.
• Achieved effective performance in unsupervised SISR tasks.
• Validated the approach on ideal, realistic, and real-world SISR scenarios.

Conclusions:

• The proposed Bayesian framework effectively models natural image statistics for improved SISR.
• The method offers robust generalizability and is suitable for unsupervised learning scenarios.
• This work advances deep learning-based image restoration by integrating principled Bayesian modeling.