Re-Boosting Self-Collaboration Parallel Prompt GAN for Unsupervised Image Restoration.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning excels in image restoration but requires large paired datasets, which are difficult to obtain in real-world scenarios.
• Unsupervised generative adversarial network (GAN)-based methods offer a solution but often struggle to outperform existing frameworks without increased complexity.
• Existing unsupervised methods face limitations in performance and computational efficiency for image restoration tasks.

Purpose of the Study:

• To propose a novel self-collaboration (SC) strategy to improve existing image restoration models.
• To enhance unsupervised image restoration performance without additional parameters or inference complexity.
• To develop a re-boosting module (Reb-SC) that integrates self-ensemble (SE) benefits without increasing inference time.

Main Methods:

• Introduced a self-collaboration (SC) strategy utilizing a prompt learning (PL) module and a restorer (Res) for iterative improvement.
• The SC strategy feeds information from previous stages to guide subsequent ones, enhancing pseudo-degraded/clean image pair generation.
• Developed a re-boosting module (Reb-SC) to incorporate self-ensemble (SE) advantages into SC without impacting inference speed.

Main Results:

• The SC strategy improved restorer performance by over 1.5dB without increasing computational complexity.
• The Reb-SC approach further boosted performance by approximately 0.3dB by integrating SE into SC.
• The proposed framework demonstrated superior performance compared to state-of-the-art unsupervised image restoration methods.

Conclusions:

• The self-collaboration strategy offers a significant advancement in unsupervised image restoration.
• The proposed methods achieve substantial performance gains with no added inference cost.
• The publicly available code and models facilitate further research and application in image restoration.