DICNO: A Generalizable Digital Image Correlation Neural Operator

Digital image correlation (DIC) is an optical metrology for dense displacement and strain fields measurement by tracking speckle patterns sprayed on the objects. As its name indicates, however, DIC relies on local image feature correlations and therefore faces challenges such as high time complexity and difficulties in computing large deformation fields. Although recent deep learning (DL)-based DIC methods have significantly accelerated the inference speed, enabling real-time measurements, these models often lack generalization to unseen deformation patterns and to varying resolutions due to their training on single deformation patterns. In addition, most DL approaches train two separate networks for the displacement and strain predictions, respectively, failing to leverage the inherent relationship between them. In this article, we introduce a novel neural operator method for DIC, digital image correlation neural operator (DICNO), that learns an operator directly mapping the latent features of two consecutive images to the underlying displacement and strain fields. By incorporating a multiresolution training strategy and continuous feature mapping, DICNO can efficiently handle various image resolutions and gain the scale generalization capability. Extensive experiments show that DICNO outperforms existing models across multiple datasets and complex scenarios, with substantial advantages in both precision and inference speed.