Variational Deep Image Restoration.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Previous convolutional neural network (CNN) methods for image restoration often assumed known degradation models or trained on diverse degradations without explicit handling of complex conditional distributions.
• Learning the conditional distribution of high-quality images from diversely degraded ones is challenging for single CNNs.
• Existing approaches sometimes incorporate additional prior information for CNN training.

Purpose of the Study:

• To propose a novel variational inference framework for image restoration.
• To develop a corresponding convolutional neural network (CNN) architecture compatible with the framework.
• To improve image restoration performance by reformulating the objective from a Bayesian perspective.

Main Methods:

• Developed a new variational inference framework that decomposes complex posterior inference into manageable sub-problems, akin to a divide-and-conquer strategy.
• Designed a convolutional neural network (CNN) structure tailored to solve image restoration problems within the proposed variational inference framework.
• Focused on reformulating the restoration objective from a Bayesian viewpoint rather than solely on network architecture or training strategies.

Main Results:

• The proposed framework significantly enhances performance across multiple image restoration tasks.
• Achieved state-of-the-art results in Gaussian denoising and real-world noise reduction.
• Demonstrated superior performance in blind image super-resolution and JPEG compression artifact reduction.

Conclusions:

• The novel variational inference framework and associated CNN offer a more effective approach to image restoration.
• The divide-and-conquer strategy derived from Bayesian reformulation leads to improved restoration quality.
• The method sets a new benchmark for various image restoration applications, including denoising, super-resolution, and artifact removal.