Interpretable DIC measurement neural network based on IC-GN algorithm framework

The digital image correlation (DIC) method is a non-contact optical measurement technique widely used to measure displacement and deformation of object surfaces due to its high precision and resolution characteristics. The traditional DIC algorithm ensures its measurement accuracy by using subset matching and iterative optimization, but it results in low measurement efficiency, difficult parameter selection, and subset smoothing effects. Recently, deep learning based DIC (DL-DIC) algorithms have achieved end-to-end output of deformation fields, avoiding the aforementioned drawbacks of traditional DIC algorithms. However, the accuracy of existing DL-DIC algorithms is still significantly lower than traditional DIC algorithms, and due to the non-interpretability of deep learning, DL-DIC algorithms have not yet been applied in the industry. To overcome the above problems, based on the traditional inverse composition-Gaussian Newton (IC-GN) DIC algorithm framework, this paper uses deep learning instead of the traditional Gaussian Newton method to obtain the incremental shape function, and iteratively optimizes the shape function using a recursive network structure to achieve high-precision deformation measurement. By this, our method not only has the advantages of high accuracy and interpretability of the IC-GN algorithm, but also has the advantages of high efficiency, simple parameter selection of DL-DIC algorithms. In the simulation experiment, our method reduced the root mean square error (RMSE) and mean absolute error (MAE) by more than 10% compared to the traditional DIC algorithm. In real material tensile tests, our method reduced the RMSE and MAE by more than 20% compared to existing DL-DIC methods.