Learn From Unpaired Data for Image Restoration: A Variational Bayes Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Generative Models

Background:

• Collecting paired training data for machine learning tasks is challenging and often impractical.
• Existing methods for unpaired data often struggle to effectively model the relationship between corrupted and clean data.

Purpose of the Study:

• To propose LUD-VAE, a deep generative method that learns joint probability density functions from marginal distributions.
• To enable the generation of synthetic training data from unpaired samples for various image enhancement tasks.

Main Methods:

• Developed a probabilistic graphical model where clean and corrupted data domains are conditionally independent.
• Utilized variational inference to maximize the evidence lower bound (ELBO) for estimating the joint probability density function.
• Demonstrated ELBO computability without paired samples under the inference invariant assumption.

Main Results:

• The proposed LUD-VAE method successfully generates synthetic training data from unpaired samples.
• Applied LUD-VAE to image denoising, super-resolution, and low-light image enhancement tasks.
• Experimental results show significant advantages of LUD-VAE over existing approaches.

Conclusions:

• LUD-VAE provides a robust mathematical framework for learning from unpaired data.
• The method effectively addresses the challenge of data scarcity in practical machine learning applications.
• LUD-VAE demonstrates superior performance in diverse image enhancement tasks.