Polarimetric image recovery method with domain-adversarial learning for underwater imaging

Area of Science:

• Computer Vision
• Image Processing
• Optics

Background:

• Underwater images suffer degradation (color shift, low contrast) due to scattering and absorption.
• Existing methods struggle with domain generalization across varied water types.
• Polarization information is crucial for improving underwater image quality.

Purpose of the Study:

• To develop a robust method for recovering degraded underwater color images.
• To address the challenge of domain generalization in underwater image restoration.
• To leverage polarization information for enhanced image recovery.

Main Methods:

• Collected the richest polarization color image dataset across different water types.
• Proposed UPD-Net, a neural network employing domain-adversarial learning.
• Integrated a water-type classifier and generative-adversarial learning for image recovery.

Main Results:

• Achieved state-of-the-art performance in visual effects and quantitative metrics.
• Demonstrated strong recovery ability in both seen and unseen underwater environments, including turbid water.
• Successfully recovered clear color images and Degree of Linear Polarization (DoLP) images.

Conclusions:

• The proposed UPD-Net effectively restores degraded underwater images across diverse water conditions.
• Domain-adversarial learning enables robust generalization for underwater image restoration.
• The method shows significant potential for real-world underwater imaging and recognition applications.