Unsupervised deep learning for 3D reconstruction with dual-frequency fringe projection profilometry

Area of Science:

• Computer Vision
• Metrology
• Machine Learning

Background:

• Fringe projection profilometry (FPP) is a key technique for high-speed, high-accuracy 3D reconstruction in industry.
• Deep learning has enhanced FPP, but requires extensive, time-consuming ground truth data for training.
• Existing methods face challenges with data annotation labor and computational resources.

Purpose of the Study:

• To develop an unsupervised convolutional neural network (CNN) for 3D reconstruction using FPP.
• To overcome the limitations of supervised learning by eliminating the need for ground truth 3D data.
• To improve the efficiency and robustness of FPP-based 3D reconstruction.

Main Methods:

• Designed an unsupervised CNN model utilizing dual-frequency fringe images.
• Introduced a fringe reprojection model for unsupervised network training.
• The network takes two fringe images and outputs a height map without ground truth data.

Main Results:

• Achieved competitive reconstruction accuracy compared to supervised methods.
• Demonstrated excellent anti-noise and generalization capabilities.
• Significantly reduced dataset generation/labeling time (by 1/6) and storage space (by 1/10).

Conclusions:

• The proposed unsupervised FPP method offers a viable alternative to supervised approaches.
• This technique reduces the burden of data annotation and storage for 3D reconstruction.
• The model provides efficient, accurate, and robust 3D reconstruction solutions.