Deep learning-based polarization 3D imaging method for underwater targets.

Area of Science:

• Optics and Photonics
• Computer Vision
• Robotics

Background:

• Underwater imaging quality is degraded by light scattering and absorption, hindering optical 3D reconstruction.
• Polarization imaging mitigates scattering effects and provides data for surface normal estimation and 3D shape reconstruction.
• Existing methods lack sufficient datasets and robust techniques for accurate underwater 3D shape reconstruction.

Purpose of the Study:

• To develop a learning-based method for accurate 3D shape reconstruction of underwater targets using polarization imaging.
• To create a comprehensive dataset for underwater polarization 3D imaging under simulated Jerlov Type I water conditions.
• To enhance the accuracy and detail of underwater 3D imaging by addressing scattering and azimuthal ambiguity.

Main Methods:

• A data acquisition system was developed to simulate underwater conditions and collect polarized images with ground truth surface normals.
• An Attention U²Net-based network framework was proposed for 3D reconstruction of underwater polarized images.
• An effective polarization representation was incorporated to resolve azimuthal ambiguity and capture detailed texture information.

Main Results:

• The proposed method effectively addresses azimuthal ambiguity, a common challenge in polarization-based reconstruction.
• Texture loss during the 3D reconstruction process was significantly reduced.
• Accuracy of surface normal estimation and overall 3D imaging was improved compared to existing techniques.

Conclusions:

• The developed learning-based shape from polarization method offers superior performance for underwater 3D imaging.
• The created dataset and proposed network framework advance the field of underwater optical 3D reconstruction.
• This approach holds promise for various applications requiring high-precision underwater 3D imaging.