Triply Complementary Priors for Image Restoration.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Deep learning models excel in image restoration (IR) but require extensive, similar training data.
• Shallow, unsupervised methods leverage image self-similarity but suffer from artifacts and slow speeds.
• Current IR methods often show limited performance and generalizability when using either deep or shallow approaches exclusively.

Purpose of the Study:

• To propose a novel joint low-rank and deep (LRD) image model for image restoration.
• To introduce a hybrid plug-and-play (H-PnP) framework integrating complementary image priors.
• To develop an effective algorithm for solving H-PnP based IR problems.

Main Methods:

• Developed a joint low-rank and deep (LRD) image model incorporating internal/external, shallow/deep, and non-local/local priors.
• Proposed a hybrid plug-and-play (H-PnP) framework leveraging the LRD model for diverse IR tasks.
• Designed a simple yet effective algorithm to address the H-PnP based IR problems.

Main Results:

• The H-PnP algorithm demonstrated favorable performance in image deblurring, compressive sensing (CS), and deblocking.
• Experimental results showed superior objective and visual perception compared to existing IR methods.
• The proposed model effectively combines the strengths of both shallow and deep learning approaches.

Conclusions:

• The proposed LRD model and H-PnP framework offer a powerful and generalizable solution for image restoration.
• This hybrid approach effectively overcomes the limitations of using deep or shallow methods in isolation.
• The developed algorithm provides a computationally efficient method for achieving high-quality image restoration.