Transformer-enhanced end-to-end models for accurate displacement and strain fields in digital image correlation

This paper presents two end-to-end digital image correlation (DIC) models—D-ST and S-ST—that leverage the Swin Transformer architecture to accurately predict full-field displacement and strain distributions. Unlike conventional DIC methods and existing CNN-based approaches, our models integrate local and global information via window-based and shifted window-based multi-head self-attention mechanisms, enabling robust and precise measurement of high-frequency deformation features. Utilizing a U-Net-like encoder-decoder framework with a multiscale feature fusion strategy, the proposed models address longstanding challenges in capturing complex strain gradients and nonlinear deformation patterns. A custom synthetic dataset, generated using B-spline finite element methods, ensures robust training and improved generalization under diverse and noisy conditions. Experimental results on both synthetic benchmarks and real-world tests highlight that D-ST and S-ST substantially outperform traditional correlation techniques and prior deep learning models, delivering stable, high-fidelity displacement and strain predictions. The approach paves the way for advancing DIC technology, facilitating higher resolution, improved accuracy, and broad applicability in material testing and structural health monitoring.