U²PNet: An Unsupervised Underwater Image-Restoration Network Using Polarization.

Area of Science:

• Computer Vision
• Optical Engineering
• Robotics

Background:

• Underwater imaging is challenging due to scattering and absorption, degrading image quality.
• Existing polarization-based restoration methods often require paired datasets or specific scene cues, limiting practical use.

Purpose of the Study:

• To develop a novel unsupervised network for underwater image restoration using polarization.
• To improve signal-to-noise ratio and overall image quality in challenging underwater environments.

Main Methods:

• Proposed U2PNet, an unsupervised network utilizing polarization information.
• Designed T-net and B∞-net subnetworks for transmission map and background light estimation.
• Developed unique non-reference loss functions for effective restoration.

Main Results:

• Achieved state-of-the-art performance on both simulated and real underwater polarization images.
• Demonstrated effective restoration without requiring pretraining datasets or specific cues.
• Validated on datasets from simulated environments and a real underwater robotic platform.

Conclusions:

• U2PNet offers a practical and effective solution for unsupervised underwater image restoration.
• The method's ability to work with single mosaicked polarized images broadens its applicability.
• The released code and datasets will facilitate further research in underwater imaging.