Time-Distributed Framework for 3D Reconstruction Integrating Fringe Projection with Deep Learning.

Area of Science:

• Computer Vision
• Optical Metrology
• Machine Learning

Background:

• Structured light techniques combined with deep learning offer high precision for 3D shape reconstruction.
• Existing methods often process data in the spatial domain, limiting applications with dynamic scenes.

Purpose of the Study:

• To propose a novel time-distributed deep learning approach for temporal structured-light 3D shape reconstruction.
• To enhance the accuracy and practicality of 3D shape reconstruction for dynamic scenarios.

Main Methods:

• Utilized an autoencoder network and a time-distributed wrapper for processing temporal fringe patterns.
• Employed fringe projection profilometry (FPP) for ground truth generation and process depiction.
• Converted fringe patterns into arctangent function numerators and denominators.

Main Results:

• The time-distributed 3D reconstruction achieved comparable results to dual-frequency datasets (p = 0.014).
• Demonstrated significantly higher accuracy than triple-frequency datasets (p = 1.029 × 10-9) via statistical tests.
• The single-network training approach proved practical for diverse applications.

Conclusions:

• The proposed time-distributed deep learning method offers a robust and accurate solution for temporal structured-light 3D shape reconstruction.
• This approach is highly practical for both scientific research and industrial applications due to its simplified implementation.
• The technique shows significant potential for advancing dynamic 3D measurement capabilities.