A monocular vision-based deep learning approach for 3D full-field displacement monitoring of planar structures

Accurate full-field three-dimensional (3D) displacement measurement is essential for understanding the dynamic behavior of planar structural components. Conventional 3D digital image correlation (3D-DIC) systems rely on stereo vision and precise calibration, resulting in increased system complexity and limited practical applicability. To address these limitations, this study proposes a deep learning-based monocular framework for 3D displacement estimation that explicitly accounts for the coupling between in-plane and out-of-plane motions. A multi-scale feature reconstruction network is first employed to estimate an apparent displacement field from image pairs. This is followed by a local affine modeling-based optical expansion network to infer out-of-plane motion. The estimated out-of-plane displacement is then used to residually compensate the apparent displacement, enabling accurate recovery of the full-field 3D displacement. To support network training and evaluation, a synthetic data generation pipeline combining deformable mesh modeling and physically based rendering is developed. Experimental validation on a plate–spring impact test demonstrates strong agreement with a stereo 3D-DIC system, with correlation coefficients exceeding 0.97 and MAEs of 0.0948, 0.1062, and 0.1689 mm in the three displacement directions. These results indicate that the proposed framework provides a reliable and cost-effective solution for full-field 3D displacement measurement in structural dynamics monitoring.