Optical-numerical method based on a convolutional neural network for full-field subpixel displacement measurements.

Area of Science:

• Optics and image processing
• Computational mechanics
• Machine learning applications

Background:

• Subpixel displacement estimation is crucial for analyzing deformations in digital images.
• Digital Image Correlation (DIC) is a common but computationally intensive method.
• Existing DIC algorithms face challenges with processing continuous image sequences.

Purpose of the Study:

• To propose a novel, efficient, and robust method for subpixel displacement estimation.
• To simplify the measurement scheme compared to traditional DIC.
• To explore the application of convolutional neural networks (CNNs) in displacement measurement.

Main Methods:

• A convolutional neural network with transfer learning (CNN-SDM) was developed.
• The CNN-SDM method compares images before and after deformation using speckle patterns.
• A coarse-to-fine estimation strategy was implemented using a series of two CNNs.

Main Results:

• The proposed CNN-SDM method demonstrated feasibility and effectiveness in experiments.
• The method achieved high efficiency, robustness, and a simple structure with few parameters.
• Simulated and real experimental results validated the CNN-SDM approach.

Conclusions:

• The CNN-SDM method offers a promising alternative for subpixel displacement measurement.
• This approach significantly reduces computational cost compared to traditional DIC.
• The method's efficiency and robustness make it suitable for various optical and image processing applications.