Deep Convolutional Neural Network Phase Unwrapping for Fringe Projection 3D Imaging

Area of Science:

• Computer Vision
• 3D Imaging
• Machine Learning

Background:

• Phase unwrapping is critical for accurate 3D reconstruction in fringe projection systems.
• Existing methods struggle with noise and large data dimensions inherent in fringe projection 3D imaging.

Purpose of the Study:

• To introduce a novel neural network for precise phase unwrapping tailored to fringe projection 3D imaging.
• To overcome limitations of current phase unwrapping techniques, particularly in noisy and large-scale datasets.

Main Methods:

• A two-step neural network training approach is proposed, utilizing the same network architecture for both steps.
• Network I is trained to identify four key features within the wrapped phase.
• Network II is trained to segment and label the wrapped phase based on features identified by Network I.

Main Results:

• The proposed neural network effectively handles wrapped phases with dimensions larger than the training data.
• Accurate phase unwrapping is achieved even in the presence of significant Gaussian noise.
• The network demonstrates robust performance when trained on simulation data and applied to experimental data.

Conclusions:

• The novel two-step neural network offers a significant advancement in phase unwrapping for fringe projection 3D imaging.
• This method provides a robust solution for accurate 3D reconstruction in challenging imaging conditions.
• The approach is validated through successful application to both simulated and real-world experimental data.