Digital image correlation based on convolutional neural networks

As an indispensable non-destructive testing technique, digital image correlation (DIC) has been increasingly ap-plied to various engineering areas concerning deformation characterization. Inspired by artificial intelligence -related technologies, we here develop a new convolutional neural network-based theoretical framework for DIC analyses, hereafter called DIC-Net. A pyramidal structure is designed to ensure robustness and reliability of mea-surement results. Simultaneously, the second-order shape function is adopted to create training dataset, making the DIC-Net more suitable for solving complex deformation fields. Different from conventional DIC algorithms, the developed DIC-Net does not require specific correlation criterion, nor is it necessary to perform numerical iterative computations, which greatly enhances the efficiency of correlation calculations. The proposed DIC-Net not only provides an alternative approach to achieve accurate, precise and reliable deformation measurements, but also paves the way for developing high-efficiency DIC with real-time processing capabilities.

Automated summary

In this study, a new theoretical framework for DIC analysis, called DIC-Net, based on convolutional neural networks is developed. The proposed DIC-Net utilizes a pyramidal structure and second-order shape functions to improve measurement robustness and reliability, as well as computational efficiency. DIC-Net offers an alternative approach to achieve accurate and reliable deformation measurements and has the potential for high-efficiency real-time DIC processing capabilities.