Nonlinear optimization DIC method inspired by unsupervised learning for high order displacement measurement

Digital Image Correlation (DIC) is the most widely used non-contact full-field strain measurement method, but due to the influence of subset calculations in principle, measuring complex deformation with high strain gradients within the subset remains a challenge. DIC based on neural networks performs well in complex deformation measurement scenarios, but its measurement accuracy and generalization ability depend on extensive training sets. This paper proposes a nonlinear optimization DIC method inspired by unsupervised learning that requires only a single pair of images. Unlike traditional unsupervised algorithms of optical flow estimation or Particle Image Velocimetry (PIV), the shape function in traditional DIC is incorporated into the loss function. This allows the use of only the reference image and the deformed image as input dataset, and the correct displacement field can be output after parameter optimization through training. This method does not require additional training sets to measure high-order strain fields like traditional neural networks-based DIC. The proposed method has demonstrated promising results in simulation experiments for high-order strain measurement and does not require additional datasets of the same type. In the measurement of the star-shaped displacement fields, its accuracy is better than the traditional DIC based on the subset, and it maintains convergence even when the strain is large, a situation where traditional DIC fails to converge. In addition, compared with the network obtained by supervised training as a pre-trained model, the method proposed in this paper can achieve superior results. The methods of the article can theoretically be applied to the result optimization of other networks beyond those discussed in this paper.