Deep 3D-DIC using a coarse-to-fine network for robust and accurate 3D shape and displacement measurements

Deep learning has become an attractive tool for addressing the limitations of traditional digital image correlation (DIC). However, extending learning-based DIC methods to three-dimensional (3D-DIC) measurements is challenging due to the limited displacement estimation range, which cannot handle the large displacements caused by stereo-matching disparities. Besides, most of the existing learning-based DIC architectures lack prior information to guide displacement estimation, resulting in insufficient accuracy. To solve these problems, we proposed a learning-based 3D-DIC (i.e., Deep 3D-DIC) using a coarse-to-fine network called G-RAFT for large and accurate image displacement estimation. Specifically, the large displacement estimation network GMA is adopted to calculate the large coarse displacement field, which is further warped on the deformed image to eliminate the main displacement component. The residual small deformation between the reference image and the warped image is further extracted using the recently proposed RAFT-DIC with high accuracy. By subtracting small displacement from large displacement, the refined displacement field is obtained. In contrast to standard subset-based 3D-DIC, Deep 3D-DIC achieves full-automatic pixel-wise 3D shape and displacement reconstruction without manual parameter input. Experimental results demonstrate that Deep 3D-DIC achieves accuracy comparable to subset-based 3D-DIC, with strong generalization ability and remarkable advantages in scenarios with complex surfaces.