Deep image correlation (DIC) is a powerful image analysis-based tool for extracting motion, full-field shape and deformation information in different material systems. It has become the industry standard for accurately measuring 2D and 3D displacement and strain fields in dynamic material testing. Compared with conventional techniques like strain gauge measurement, DIC offers numerous benefits, including the ability to perform full-field measurement with direct contact with the samples, easy to set up experimental solutions with less strict conditions and improved measurement accuracy and robustness.
In the last decades, DIC algorithms have undergone tremendous development to improve better accuracy and computational efficiency. Most algorithms are divided into two categories: local subset-based and global continuum methods. The former relies on interpolation using correlation matrix and gray-scale pixel values, while the latter uses a set of solved finite element methods of shape functions to represent the displacement field image. Recently, deep learning has received great attention in numerous computer vision tasks like object identification and image classification with promising successes.
Though incorporating DIC with deep learning for material characterization seems a promising concept, substantial success is yet to be achieved due to three key challenges. First, there is no reliable demonstration of full-field strain field prediction using deep learning techniques. Second, existing deep-learning-based techniques have not exhibited remarkable performance superiority over traditional DIC. Lastly, there are limited attempts to compare the accuracy and overall performance of traditional and deep-learning base DIC. Although traditional DIC is practically viable for general tensile testing cases, it has several deficiencies, such as lengthy computation, low-resolution output and unstable prediction at large deformation.
To overcome these challenges, Dr. Ru Yang and Professor Ping Guo from Northwestern University in collaboration with Dr. Yang Li and Dr. Danielle Zeng from Ford Motor Company developed a novel end-to-end deep learning-based DIC approach for (named Deep DIC) for robust and accurate measurement and prediction of both full-field displacement and strain fields. In their approach, two separate convolutional neural networks (CNNs): DisplacementNet and StrainNet were designed using a modified encoder-decoder structure. They worked independently to achieve end-to-end displacement and strain field predictions as well as tracking large deformations. Their work is currently published in the Journal of Materials Processing Technology.
The authors showed that the Deep DIC outperformed traditional DIC on real experimental data despite being trained only on synthetic datasets. DisplacementNet was used to predict the displacement field and adaptively track large deformations by updating the region of interest, while StrainNet allowed a direct strain field prediction from the image input without depending on the displacement prediction or spatial derivative-based calculations. Generally, deep DIC improved overall prediction accuracy and required no iteration or interpolation to solve strain and displacement fields from image pairs.
A new dataset generation method was designed to synthesize comprehensive and realistic datasets for training the proposed model. The real-life performance of Deep DIC was evaluated and validated. The new dataset generation method increased the adaptability and robustness of the Deep DIC by rendering the speckle patterns with various qualities to prescribe a wide range of random deformation and motion. The Deep DIC achieved comparable and consistent predictions as those of commercial DIC software. It, however, outperformed the commercial software in terms of computational time, adaptability and robustness even at large deformation and different pattern qualities.
In summary, the development of Deep DIC for robust and accurate end-to-end measurement and prediction of displacement and strain fields for different material testing was reported. It utilizes separate CNNs working independently to directly achieve strain predictions, thereby avoiding noise and potential errors induced by the displacement field. This approach reduced the prediction cost while preserving the spatial resolution without post-filtering. In a statement to Advances in Engineering, Professor Ping Guo explained their study provided more insights for advancing the application of DIC for material characterization and tensile testing.
NOTE: The code, model, and dataset for Deep DIC are released on
Yang, R., Li, Y., Zeng, D., & Guo, P. (2022). Deep DIC: Deep learning-based digital image correlation for end-to-end displacement and strain measurement. Journal of Materials Processing Technology, 302, 117474.