Physics-based supervised learning method for high dynamic range 3D measurement with high fidelity.

Area of Science:

• Optics and Photonics
• Computer Vision
• Metrology

Background:

• High dynamic range (HDR) 3D measurement is crucial but challenging.
• Current deep learning methods struggle with complex reflectivity and illumination, leading to poor convergence and robustness.
• Existing unsupervised methods lack detail recovery for complex surfaces.

Purpose of the Study:

• To develop a robust and accurate HDR 3D measurement method.
• To improve network convergence and generalization for diverse surface properties and lighting.
• To achieve high-fidelity recovery of fine surface details.

Main Methods:

• A physics-based supervised learning approach was developed.
• A novel sinusoidal-component-to-sinusoidal-component mapping paradigm was introduced, incorporating a physical model for phase retrieval.
• The method eliminates scale differences in fringe intensity across various illumination conditions.

Main Results:

• The proposed method significantly enhances network convergence and generalization compared to conventional supervised methods.
• It outperforms unsupervised methods in recovering complex surfaces with greater detail.
• Experiments demonstrated high-quality phase recovery (STD error ~0.03 rad) across diverse materials (diffuse, metal, hybrid) and lighting, validating superior generalization.

Conclusions:

• The physics-based supervised learning method offers a robust solution for HDR 3D measurement.
• The novel mapping paradigm effectively handles complex reflectivity and illumination variations.
• The method demonstrates high fidelity in recovering intricate surface details and strong generalization capabilities.