TY - JOUR
T1 - Improving Deep Learning‑Based Digital Image Correlation with Zero Mean Normalized Cross-Correlation
AU - Gao, J.
AU - Wang, G.
AU - Wang, P.
AU - Chen, H.
AU - Yang, H.
N1 - Publisher Copyright:
© Society for Experimental Mechanics 2025.
PY - 2026/2
Y1 - 2026/2
N2 - Background: In recent years, deep learning-based digital image correlation (DIC) methods have been widely adopted. However, fully data-driven algorithms are often constrained by the quality and diversity of training datasets. Moreover, conventional convolutional neural network (CNN) approaches for feature extraction exhibit severely reduced accuracy when processing speckle images with poor-quality patterns or significant noise. Objective: To enhance the generalization ability and robustness to noise of deep learning-based digital image correlation methods when processing images with varying speckle patterns, while ensuring high computational efficiency. Methods: This paper proposes a novel approach for calculating the deformation field of natural textured speckle images using a normalized cross-correlation layer. Building on the conventional StrainNet CNN, the method incorporates the zero-mean normalized cross-correlation criterion and integrates subset matching algorithms from traditional DIC methods. By adjusting the parameters of the normalized cross-correlation layer, the proposed method can adapt to different speckle patterns and deformation modes, enabling robust correlations between reference and current images. Furthermore, the output channels are unified at the output end and can be automatically adjusted according to the parameters of the correlation layer. The adjusted correlation calculation results are directly fed into the multi-level feature extraction layer, thereby optimizing the network architecture and enhancing network flexibility. Results: Through numerical and real experiments, the proposed method achieves higher accuracy with a smaller dataset when applied to images with randomly generated high-quality speckle and natural textures. Compared to traditional DIC algorithms, the method improves computational efficiency by about four orders of magnitude while maintaining the same level of accuracy. For experimental images with poor speckle quality and high noise level, the method reduces the mean absolute error by over 0.11 pixels compared to StrainNet, demonstrating superior generalizability to various speckle patterns. Conclusion: The results demonstrate that the proposed deep learning deformation calculation method, which integrates data-driven approaches and a correlation layer, significantly enhances computational efficiency and generalization capability. This method not only adapts more effectively to low-quality natural textured speckle patterns, but also exhibits greater robustness when handling images affected by unknown noise.
AB - Background: In recent years, deep learning-based digital image correlation (DIC) methods have been widely adopted. However, fully data-driven algorithms are often constrained by the quality and diversity of training datasets. Moreover, conventional convolutional neural network (CNN) approaches for feature extraction exhibit severely reduced accuracy when processing speckle images with poor-quality patterns or significant noise. Objective: To enhance the generalization ability and robustness to noise of deep learning-based digital image correlation methods when processing images with varying speckle patterns, while ensuring high computational efficiency. Methods: This paper proposes a novel approach for calculating the deformation field of natural textured speckle images using a normalized cross-correlation layer. Building on the conventional StrainNet CNN, the method incorporates the zero-mean normalized cross-correlation criterion and integrates subset matching algorithms from traditional DIC methods. By adjusting the parameters of the normalized cross-correlation layer, the proposed method can adapt to different speckle patterns and deformation modes, enabling robust correlations between reference and current images. Furthermore, the output channels are unified at the output end and can be automatically adjusted according to the parameters of the correlation layer. The adjusted correlation calculation results are directly fed into the multi-level feature extraction layer, thereby optimizing the network architecture and enhancing network flexibility. Results: Through numerical and real experiments, the proposed method achieves higher accuracy with a smaller dataset when applied to images with randomly generated high-quality speckle and natural textures. Compared to traditional DIC algorithms, the method improves computational efficiency by about four orders of magnitude while maintaining the same level of accuracy. For experimental images with poor speckle quality and high noise level, the method reduces the mean absolute error by over 0.11 pixels compared to StrainNet, demonstrating superior generalizability to various speckle patterns. Conclusion: The results demonstrate that the proposed deep learning deformation calculation method, which integrates data-driven approaches and a correlation layer, significantly enhances computational efficiency and generalization capability. This method not only adapts more effectively to low-quality natural textured speckle patterns, but also exhibits greater robustness when handling images affected by unknown noise.
KW - Deep learning
KW - Digital image correlation
KW - Natural textured speckle patterns
KW - Normalized cross-correlation layer
UR - https://www.scopus.com/pages/publications/105018321969
U2 - 10.1007/s11340-025-01231-9
DO - 10.1007/s11340-025-01231-9
M3 - Article
AN - SCOPUS:105018321969
SN - 0014-4851
VL - 66
SP - 237
EP - 252
JO - Experimental Mechanics
JF - Experimental Mechanics
IS - 2
ER -