Unsupervised learning polarimetric underwater image recovery under nonuniform optical fields

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Turbid media significantly degrade image quality.
• Polarization imaging offers a solution for clear imaging in scattering environments.
• Existing methods may struggle with nonuniform optical fields and require extensive datasets.

Purpose of the Study:

• To develop an improved polarization imaging method for turbid media.
• To address challenges posed by nonuniform optical fields.
• To enhance real-time performance and reduce dataset preparation costs.

Main Methods:

• Combined unsupervised learning with polarization imaging theory.
• Designed an end-to-end unsupervised generative network to inpaint spatially variable background light.
• Utilized adversarial loss with a discriminative network for improved performance.
• Estimated polarization parameters using the angle of polarization.

Main Results:

• Demonstrated the effectiveness and generalization ability of the proposed method.
• Achieved superior real-time performance compared to existing methods.
• Required a lower cost for training dataset preparation.

Conclusions:

• The proposed unsupervised learning-based polarization imaging method effectively enhances image quality in turbid media.
• The approach is robust across various nonuniform optical fields.
• Offers a practical and cost-effective solution for clear imaging applications.