Unsupervised-learning-based calibration method in microscopic fringe projection profilometry

Area of Science:

• Metrology and Measurement Science
• Computer Vision and Image Processing
• Machine Learning Applications

Background:

• Microscopic fringe projection profilometry (MFPP) is crucial for 3D measurement, but its precision relies heavily on accurate calibration.
• Shallow depth of field in MFPP often leads to low-quality target images, causing inaccurate feature detection and calibration parameter estimation.
• Existing calibration methods struggle with image degradation, limiting overall measurement accuracy.

Purpose of the Study:

• To develop an unsupervised learning-based calibration method for MFPP that is robust to image defocus and noise.
• To enhance image quality and improve calibration accuracy in MFPP systems.
• To achieve superior performance in both calibration accuracy and measurement precision.

Main Methods:

• An unsupervised image deblurring network (UIDNet) was developed to restore sharp target images from degraded inputs.
• The UIDNet utilizes a multi-quality target dataset, avoiding paired image capture or simulation for more accurate feature learning.
• Multi-perceptual loss and Fourier frequency loss were integrated into UIDNet for enhanced training.
• A robust calibration compensation strategy employing 2D discrete Fourier transform was implemented to assess image quality and refine feature center detection.

Main Results:

• The proposed unsupervised learning approach significantly improves image quality, recovering sharp target images from defocused and noisy inputs.
• The method enhances the accuracy of reference feature center detection, crucial for precise calibration.
• Experimental results demonstrate superior performance in calibration accuracy compared to traditional methods.
• The improved calibration directly translates to enhanced measurement precision in MFPP systems.

Conclusions:

• The unsupervised learning-based calibration method effectively addresses image quality issues in MFPP, particularly defocus and noise.
• The integration of UIDNet with multi-perceptual and Fourier frequency losses optimizes the deblurring process.
• The developed strategy provides a robust solution for accurate feature detection and calibration, leading to higher measurement precision.
• This approach offers a significant advancement for reliable and accurate 3D measurements using MFPP technology.