Polarimetric binocular three-dimensional imaging in turbid water with multi-feature self-supervised learning

Polarization imaging provides significant advantages in underwater environments. However, existing polarization underwater imaging methods primarily focus on leveraging polarization information to suppress the scattering effect to achieve the clear vision, while neglecting other valuable information contained in polarization images, such as the scene depth and the polarization characteristics of the objects. This paper proposes a self-supervised three-dimensional underwater imaging method based on a polarization binocular imager. In addition to improving image quality in turbid water based on polarization imaging, the proposed method merges features from both the enhanced binocular images recovered from polarization information and the feature-rich degree of polarization images into the self-supervised framework to estimate disparities of the scene, achieving high-quality reconstruction of underwater scene depth. We then design multiple self-supervised losses that effectively integrate depth information obtained from both binocular imaging and polarization imaging to guide the learning process. Meanwhile, the proposed method can recover the polarization information of the objects in turbid water, thus enhancing the perception of target properties such as the materials of the objects. Both the simulated experiment and the real-world experiments in the sea demonstrate the effectiveness and superiority of the proposed method.