Dual Prior Learning for Blind and Blended Image Restoration

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Unsupervised image restoration often relies on Deep Image Prior (DIP) to learn image statistics.
• Images can be corrupted by multiple, unknown distortions, making restoration challenging for methods relying solely on image priors.
• Disentangling clean image signals from hybrid distortions is difficult with existing approaches.

Purpose of the Study:

• To develop a novel unsupervised image restoration method capable of handling multiple unknown distortions.
• To improve upon Deep Image Prior (DIP) by incorporating distortion priors.
• To effectively disentangle image and distortion characteristics for enhanced restoration.

Main Methods:

• Proposed Dual Prior Learning (DPL) method, which learns both image and distortion priors.
• Incorporated an explicit step to learn a blended distortion prior.
• Utilized unpaired training data with weak adversarial supervision to disentangle the two priors.

Main Results:

• Demonstrated the effectiveness of DPL in restoring images with challenging, unknown blended distortions.
• Achieved appealing performance compared to existing methods.
• Showcased the ability of DPL to handle complex degradation scenarios.

Conclusions:

• Dual Prior Learning (DPL) offers a significant advancement in unsupervised image restoration for complex distortions.
• The method successfully disentangles image and distortion priors, leading to superior restoration quality.
• DPL provides a robust framework for tackling real-world image restoration challenges.