Three-dimensional Digital Image Correlation Method Based on Physics-Informed Neural Network

<jats:p>Combined with stereo vision, digital image correlation (DIC) enables three-dimensional digital image correlation (3D-DIC) for full-field measurement of surface morphology and three-dimensional deformation. However, conventional subset-based 3D-DIC methods suffer from strong dependence on empirically selected parameters, limited capability in handling complex non-uniform deformations, and reduced measurement accuracy near irregular boundaries. In recent years, physics-informed neural network digital image correlation (PINN-DIC) has achieved promising performance in two-dimensional displacement field measurement. Nevertheless, due to the larger disparity magnitude and more complex spatial variations induced by stereo imaging, directly extending PINN-DIC to 3D-DIC still faces significant challenges, including a large optimization search space, convergence difficulties, and sensitivity to non-uniform illumination. To address these issues, this paper proposes a seed-point-scale-guided 3D-PINN-DIC method. First, sparse seed-point matching is employed to estimate disparity scale priors, which are further used to normalize the network output, thereby reducing the optimization search space and improving the convergence stability for large-scale disparity field reconstruction. Second, a convolution-kernel-based zero-mean normalized sum of squared differences (ZNSSD) criterion is constructed to enhance robustness against non-uniform illumination and brightness variations. The proposed method is validated through both numerical simulations and real experiments, and its performance is compared with that of conventional subset-based methods. The results demonstrate that the proposed approach achieves higher measurement accuracy in regions with complex deformation and disparity variations, while reconstructing smoother and more continuous disparity fields, confirming its effectiveness for high-precision three-dimensional deformation measurement under challenging conditions.</jats:p>