ECI-Net: A modular and lightweight Deep DIC network for real-time robust 2D displacement measurement

Deep learning-based Digital Image Correlation (Deep DIC) is considered a potential method to overcoming the issues of Digital Image Correlation (DIC), such as accuracy, and efficiency. However, the current Deep DICs may struggle to accommodate various types of deformations, which affects their stability. To address the issue, referring to the previous Deep DIC architecture and traditional DIC, we propose a lightweight architecture called ECI-Net, which includes three modules: extractor, correlator, and integrator. To enable the network to generalize across multiple displacement modes within a +/- 10 pixel range, an artificial speckle dataset is generated, which helps the network better adapt to small deformation scenarios encountered in real-world applications. The proposed modules' effectiveness is demonstrated through ablation experiments. The constancy of ECI-Net is proved through a test set that included various displacement modes various displacement modes and ranges. Meanwhile, ECI-Net is also validated as a stable Deep DIC in real experiments.