A lightweight deep learning model DICNet3+ for large deformation measurement in digital image correlation

Abstract

Accurate deformation measurement is essential for evaluating material performance in complex mechanical testing. Although the traditional digital image correlation method is widely used, it faces limitations, such as boundary instability and erroneous data due to speckle pattern tearing, especially in large deformation scenarios. To address these challenges, this study proposes a lightweight deep learning model DICNet3+ , which is based on a modified UNet3+ architecture incorporating depthwise separable convolutions and convolutional block attention modules. These enhancements improve feature extraction while minimizing the number of parameters, enabling accurate prediction of displacement fields in large deformation scenarios. A comprehensive dataset consisting of both real and simulated speckle patterns, and a weighted hybrid loss function that combines root mean square error and average endpoint error were developed to train and validate the model. The results demonstrated that the DICNet3+ model significantly outperformed existing deep learning-based DIC models in terms of accuracy, robustness, and generalization. Additionally, the DICNet3+ model provided reliable predictions even in regions with erroneous data or along boundaries and showed significant computational efficiency compared to ARAMIS software in compression experiments, particularly when GPU acceleration was used. This work made DICNet3+ a viable solution for large deformation measurements in engineering applications.