Digital image correlation (DIC) as a non-contact deformation measurement technique is widely used in various engineering fields. Recently, satisfactory results have been achieved by deep learning-based DIC methods, which rely on synthetic datasets for supervised training due to the difficulty in obtaining the ground truth of deformation fields. However, generalization to real scenes remains challenging due to the inherent differences between synthetic and real images and the limited variability of synthetic datasets. To solve this problem, this paper proposes an unsupervised learning DIC method, referred as UnDIC-Net. UnDIC-Net proposes a Patch-based Zero-Normalized Sum of Squared Differences (Patch-ZNSSD) similarity measure as the loss function for unsupervised training, thereby eliminating the dependency on hard-to-obtain labeled data. Meanwhile, we also collected speckle image from real scenes and built an unlabeled dataset to train UnDIC-Net. For network structure, UnDIC-Net adopts an incremental deformation estimation strategy to compute the deformation field, and interestingly, this network can be applied to large deformation measurements. Adequate experiments show that UnDIC-Net performs well in small deformation measurements, and for larger deformation measurements (>100 pixels), some traditional methods fail while UnDIC-Net still performs well. The code and data of this paper is released at: https://github.com/LianpoWang/UnDICnet-master.
Read full abstract