Polarization technology facilitates 3D imaging under turbid water

A collaborative research team led by Prof. Haofeng Hu from the School of Marine Science and Technology at Tianjin University—partnering with researchers from Northwestern Polytechnical University, Harvard Medical School and Nanyang Technological University—has developed a novel polarimetric binocular three-dimensional imaging method. Leveraging multi-feature self-supervised learning, this innovation tackles longstanding challenges in underwater imaging: it simultaneously enhances image clarity and enables high-precision depth estimation in turbid water, marking a critical advancement for marine exploration and related fields. 

This study focuses on the two core points of underwater optical imaging in turbid environments: image degradation from light scattering and the lack of depth perception. Existing underwater polarization imaging only focuses on the descattering task while ignoring depth cues and the intrinsic polarization characteristics of objects. This team first extracted multi-features in polarization images, discovering the depth cues implicit in polarization information. 

Building on this foundation, the team designed a multi-feature self-supervised depth estimation framework. This framework fuses features from enhanced restored binocular images and DoP (Degree of Polarization) images, then feeds these fused features into a disparity estimation network. This network extracts multi-scale features through dimensionality reduction in the encoder, and the decoder paired with skip connections can achieve precise disparity prediction. To eliminate reliance on ground truth depth data, the team constructed multiple self-supervised loss functions, allowing the network to learn effectively without labeled depth data. 

To validate the method comprehensively, the team conducted two types of experiments. First, they generated simulated underwater polarimetric binocular data to verify the algorithm’s depth estimation capability under known conditions. Second, they collected real-world data in marine environments using a remotely operated vehicle (ROV) equipped with the polarimetric binocular imager, testing the method’s performance in improving image quality and reconstructing 3D scenes in complex turbid waters.

In conclusion, the study achieves innovative breakthroughs in three key areas: 

① Pioneering Polarization Depth Cues: Traditional underwater polarization imaging has long been limited to image descattering, failing to fully exploit depth perception potential of polarization imaging. Through extensive experiments and theoretical analysis, this study is the first to reveal the unique value of Polarization Difference Images in underwater depth characterization. Generated by fusing recovered scene images and DoP information, it effectively overcomes depth estimation errors caused by diversities of object surface polarization characteristics. This advancement significantly enhances the ability to distinguish between near and distant objects, enabling polarization imaging to evolve from traditional "image enhancement" to "depth perception" . 

② Multi-Feature Self-Supervised Learning Framework: To solve the scarcity of labeled data for underwater depth estimation, the team built an innovative multi-feature self-supervised framework. Taking enhanced binocular images and DoP images as inputs, it  incorporates a cross-scale fusion strategy for the network, enabling multi-scale disparity prediction. With its self-supervised mechanism, the framework requires no real depth data for training, effectively solving the problem of the lack of depth data in underwater scenes. Experimental results show it outperforms traditional supervised learning methods in estimation accuracy, even without labeled data. 

③ Polarimetric-Binocular Data Acquisition and Algorithm Robustness: The team has  acquired binocular polarization data using the polarimetric binocular cameras in both laboratory environments and real sea. Extensive validations across scenarios demonstrate the method’s excellent robustness under varying water quality conditions. Notably, in extremely turbid environments, its depth estimation error is significantly lower than that of traditional binocular methods; it also effectively avoids depth missing in low-texture regions (e.g., smooth target surfaces, turbid backgrounds), greatly improving the reliability of depth estimation in complex underwater scenes .