Single-frame two-stage fringe projection profilometry based on deep learning.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Traditional fringe projection profilometry faces challenges in accurately measuring 3D information for dynamic objects with surface discontinuities.
• Existing methods often require multiple fringe patterns or struggle with complex surfaces.

Purpose of the Study:

• To develop an accurate and efficient 3D measurement technique for objects with surface discontinuities using a single fringe pattern.
• To leverage deep learning for enhanced phase prediction and unwrapping in fringe projection profilometry.

Main Methods:

• A single-frame, dual-stage fringe projection profilometry technique employing two neural networks.
• The first neural network predicts fringe patterns at various frequencies; the second predicts wrapped phase numerator and denominator.
• Integration of a multi-frequency phase unwrapping method with system calibration.
• Introduction of DARU-Net, a novel convolutional neural network based on U-Net architecture.

Main Results:

• Accurate prediction of 3D information for objects with surface height discontinuities from a single fringe pattern.
• DARU-Net demonstrated superior performance compared to U-Net and its derivatives in deep learning tasks.
• The proposed method successfully overcomes limitations of traditional profilometry for complex surfaces.

Conclusions:

• The developed single-frame, dual-stage technique offers a robust solution for 3D measurement of challenging objects.
• This approach significantly expands the application scope of fringe projection profilometry in dynamic and discontinuous scenarios.
• The integration of advanced deep learning models like DARU-Net enhances the accuracy and efficiency of 3D reconstruction.