Single-shot fringe projection profilometry based on deep learning and computer graphics.

Area of Science:

• Computer Vision
• Optical Metrology
• Machine Learning

Background:

• Deep learning has been increasingly applied to fringe projection profilometry (FPP).
• Acquiring large real-world datasets for training FPP models is challenging.
• Network design and optimization for FPP remain areas for exploration.

Purpose of the Study:

• To develop a method for conveniently generating large datasets for FPP using virtual systems.
• To design a novel loss function for accurate depth estimation from single fringe images.
• To compare the performance of different deep learning network architectures for FPP.

Main Methods:

• Construction of virtual FPP systems using graphic software.
• Analysis of key factors for virtual system realism.
• Development of a new loss function to improve depth image quality and detail restoration.
• Comparison of U-Net and pix2pix network architectures.

Main Results:

• Virtual FPP systems successfully generated diverse datasets.
• The novel loss function enhanced depth estimation accuracy and detail.
• Networks trained on virtual data demonstrated good accuracy and generalization in real experiments.
• The study provides guidance for applying deep learning to FPP.

Conclusions:

• Virtual FPP systems offer a practical solution for data generation.
• The proposed deep learning approach with the new loss function is effective for depth estimation.
• The findings support the use of simulated data for training FPP models.