Single-shot 3D measurement of highly reflective objects with deep learning.

Area of Science:

• Metrology
• Computer Vision
• Machine Learning

Background:

• Fringe projection profilometry (FPP) is crucial for industrial 3D measurement.
• Traditional FPP methods struggle with dynamic scenes and overexposure from reflective surfaces.
• Existing techniques often require multiple images, limiting real-time applications.

Purpose of the Study:

• To develop a single-shot, high dynamic range 3D measurement method for FPP.
• To address limitations of dynamic scene measurement and overexposure in industrial parts.
• To enhance the accuracy and applicability of FPP in challenging industrial environments.

Main Methods:

• A deep learning model comprising an exposure selection network (ExSNet) and a fringe analysis network (FrANet).
• ExSNet uses a self-attention mechanism to handle overexposure in highly reflective areas.
• FrANet predicts wrapped and absolute phase maps for 3D reconstruction.
• A specialized training strategy optimizes for measurement accuracy.

Main Results:

• The method accurately predicts optimal exposure time in a single shot.
• Quantitative evaluation on moving spheres with overexposure showed low prediction errors (diameter: 73µm/64µm, center distance: 49µm).
• Successful reconstruction of standard spheres across a wide exposure range was demonstrated.

Conclusions:

• The proposed deep learning approach enables accurate single-shot, high dynamic range 3D measurement using FPP.
• This method significantly improves 3D measurement capabilities for dynamic industrial parts with reflective surfaces.
• The findings offer a robust solution for challenging metrology applications in manufacturing.