StrainNet-LD: Large Displacement digital image correlation based on deep learning and displacement-field decomposition

First, a displacement-field decomposition strategy is proposed that effectively enhances the accuracy of digital image correlation (DIC) algorithms under large displacement or deformation conditions. The implementation methodologies of displacement-field decomposition under both Lagrangian and Eulerian descriptions are derived and presented. Second, displacement-field decomposition is applied to speckle image correlation algorithms based on deep learning, resolving the contradiction between registration capability of large displacement and the accuracy. A CUDA-accelerated and residual learning-based strain fitting algorithm is proposed for high-precision and real-time strain computation. This method is named "StrainNet for Large Displacement" (StrainNet-LD). StrainNet-LD achieves a calculation speed of 768 x 768 x 6 fps (768 x 768 x 3 fps with strain calculated simultaneously) on GPU with displacement MAE < 0.05 pixels and strain MAE < 0.002 epsilon. Finally, rubber tensile tests, 3D motion measurement of composite shells, and morphology reconstruction experiments of speckle-free composite repair structures are conducted to validate the algorithm's robustness and efficiency. Some vital discussions of displacement-field decomposition are given in the Appendixes. The code and dataset are available at https://github.com/GW-Wang-thu/StrainNet-LD.