Shape from polarization via a physical prior-based deep fusion network with ambiguous surface normals

Shape from polarization imaging is a passive three-dimensional imaging method with high precision and detailed reconstruction capabilities, widely used in fields such as intelligent manufacturing, surface defect detection, and medical imaging. However, existing deep learning-based shape from polarization methods have not effectively fused polarization and other physical prior information, leading to degraded reconstruction quality in complex scenes. To address this issue, we propose an innovative method that integrates polarization information with ambiguous surface normals and specular confidence information. By incorporating these additional prior features, the model's reconstruction accuracy in complex scenes is significantly improved. This research calculates the ambiguous surface normals and specular confidence information based on a physical model and introduces a novel dual-branch deep fusion network. The network efficiently extracts multi-scale feature information through a feature extraction module and effectively fuses polarization information with ambiguous surface normals and specular confidence information via a feature fusion module, enhancing the reconstruction accuracy in complex scenes. Experimental results demonstrate that the proposed method can accurately reconstruct surface normal under complex lighting and low-texture conditions, significantly improving the accuracy and robustness of shape from polarization methods. The method is expected to have broad applications in intelligent manufacturing, surface defect detection, and medical imaging. Our dataset and source code will be publicly available at https://github.com/singobl/CGA-Transformer.