Improving Deep Learning-Based Digital Image Correlation with Domain Decomposition Method

Abstract

Background

Deep learning-based digital image correlation (DL-based DIC) has gained increasing attention in the last two years. However, existing DL-based DIC algorithms are impractical because their application scenarios are mostly limited to small deformations.

Objective

To enable the use of DL-based DIC in real-world general experimental mechanics scenarios that would involve large deformations and rotations, we propose to improve DL-based DIC with the domain decomposition method (DDM).

Methods

In the improved method, the region of interest is divided into subimages, and subimages are pre-aligned using the preregistered control points to effectively eliminate the large deformation components. The residual deformations in each subimage are small and limited, which can be well extracted using existing DL-based DIC methods.

Results

Through synthesized and real-world experiments, the improved DL-based DIC method can achieve high-accuracy pixelwise matching in practical applications with strong robustness and high computational efficiency.

Conclusions

The improved DL-based DIC combines the advantages of traditional and DL-based DIC methods but overcomes the limitations, greatly improving the robustness and applicability of existing DL-based methods.