Exploring the Potential of Pooling Techniques for Universal Image Restoration

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Image restoration aims to recover clean images from degraded versions.
• Transformers have shown promise but suffer from quadratic complexity.
• Existing methods often increase complexity, necessitating simpler yet effective approaches.

Purpose of the Study:

• To propose an efficient and effective mechanism for image restoration.
• To explore the potential of pooling techniques for implicit dual-domain representation learning.
• To develop a novel pooling module that enhances performance across various restoration tasks.

Main Methods:

• Leveraging average and max pooling for implicit low- and high-frequency signal extraction.
• Utilizing lightweight learnable parameters for frequency component modulation.
• Incorporating dual-domain modulation across multiple scales and shapes within the pooling module.
• Employing intermediate high-frequency features as attention maps for edge information highlighting.

Main Results:

• Achieved state-of-the-art performance on 15 datasets for five single- and two composite-degradation image restoration tasks.
• Demonstrated effectiveness in single-degradation, composite-degradation, and all-in-one image restoration.
• Showcased favorable performance against state-of-the-art all-in-one algorithms under two settings.

Conclusions:

• The proposed pooling module offers an efficient and effective approach to image restoration.
• The method successfully addresses limitations of complex models like Transformers.
• The pooling module is versatile and performs well in diverse image restoration scenarios, including all-in-one restoration.