Convolutional neural network for estimating physical parameters from Newton's rings.

Area of Science:

• Optical metrology
• Interferometry
• Machine learning applications

Background:

• Newton's rings are a classic interferometric phenomenon used for surface analysis.
• Traditional methods for analyzing Newton's rings can be sensitive to noise and time-consuming.

Purpose of the Study:

• To apply convolutional neural networks (CNNs) for automated parameter estimation of Newton's rings.
• To evaluate the precision, noise immunity, and speed of CNN-based methods compared to traditional approaches.

Main Methods:

• Implementation of Visual Geometry Group (VGG) network and U-Net architectures for Newton's rings analysis.
• Training of CNN models to simultaneously estimate the center and curvature radius of Newton's rings.
• Testing models on noisy image data (Gaussian and salt-and-pepper noise).

Main Results:

• Both VGG and U-Net models successfully estimated Newton's rings parameters.
• The VGG model achieved estimation in 0.01 seconds for a 640x480 pixel image.
• Parameter estimation errors were less than one pixel for the center and below 0.5% for the curvature radius, even with significant noise.

Conclusions:

• Deep learning, specifically VGG and U-Net, offers a highly precise and robust method for Newton's rings analysis.
• CNN-based approaches significantly reduce analysis time and improve noise immunity over conventional techniques.
• This method enables efficient and accurate characterization of spherical surfaces using interferometric patterns.