Underwater polarization image de-scattering utilizing a physics-driven deep learning method

Abstract

The remarkable ability of polarization imaging to suppress the backscattered light makes it a highly attractive solution for various underwater applications. In recent years, emerging learning-based polarization technologies have shown significant potential for application and achieved great success. However, the majority of learning-based studies primarily employ data-driven approaches, which lack interpretability and generalizability. To address this problem, we propose a polarization de-scattering method in which the combination of an active polarization imaging model with deep learning is well executed. Firstly, the network can focus more attention on specific polarization information by applying a well-designed polarization feature-refined block. Secondly, the network directly predicts the polarization-related parameters of the active polarization imaging model, eliminating the need for prior parameters and manual estimation during its operation. Lastly, the network generates clear de-scattered images under the guidance of the model. Additionally, we design efficient loss functions to fully restore the polarization information of degraded images and further improve the recovery performance of intensity information. Several groups of experimental results demonstrate that our method outperforms other advanced methods for targets with different materials and under varying turbidity conditions.