Plug-and-Play Image Restoration With Deep Denoiser Prior

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Plug-and-play image restoration leverages denoisers as image priors for inverse problems.
• Deep convolutional neural networks (CNNs) offer powerful denoiser priors but existing methods face limitations.
• There is a need for improved denoiser priors to advance plug-and-play image restoration.

Purpose of the Study:

• To develop a benchmark deep denoiser prior for enhancing plug-and-play image restoration.
• To integrate this denoiser prior into a model-based iterative algorithm.
• To analyze the method's performance and working mechanism across various image restoration tasks.

Main Methods:

• Trained a highly flexible and effective CNN denoiser as a benchmark prior.
• Integrated the deep denoiser prior into a half quadratic splitting based iterative algorithm.
• Evaluated the approach on deblurring, super-resolution, and demosaicing tasks.

Main Results:

• The proposed method significantly outperforms state-of-the-art model-based approaches.
• Achieved competitive or superior performance compared to state-of-the-art learning-based methods.
• Demonstrated effectiveness across diverse image restoration challenges.

Conclusions:

• A benchmark deep denoiser prior effectively enhances plug-and-play image restoration.
• The proposed method offers a flexible and powerful solution for inverse problems.
• This work advances the integration of deep learning priors in model-based image restoration.