Physics-informed neural network for polarimetric underwater imaging.

Area of Science:

• Optics
• Computer Vision
• Marine Technology

Background:

• Polarization analysis effectively suppresses scattered light in turbid water for improved underwater imaging.
• Existing neural networks for polarimetric imaging often neglect physical models, limiting performance.

Purpose of the Study:

• To develop a deep neural network integrating polarimetric physical models for enhanced underwater image restoration.
• To improve the convergence and performance of polarimetric imaging networks by incorporating physical constraints.

Main Methods:

• Developed a deep neural network informed by a polarimetric physical model.
• Mathematically transformed two polarization-dependent parameters into a single parameter for easier network convergence.
• Designed a polarization perceptual loss to leverage polarization information at the feature level.

Main Results:

• The proposed network successfully learned the polarization-modulated parameter, yielding clear de-scattered images.
• Experimental results demonstrated superior image quality compared to existing methods, even in high turbidity.
• The integration of physical models and neural networks significantly improved underwater image restoration.

Conclusions:

• Combining polarimetric physical models with deep neural networks offers a powerful approach for de-scattering underwater images.
• The proposed method effectively restores target signals in turbid water by leveraging polarization information.
• This technique shows significant potential for applications requiring high-quality underwater imaging.