Transformer‐Based Automatic Tracking of Structural Crack Propagation

As aging infrastructure presents increasing safety and maintenance challenges, effective monitoring of structural defect propagation is vital. While many methods detect cracks, few can track the same crack over time under varying conditions. This study presents a modular, automated framework to track crack propagation by integrating scale‐invariant feature transform (SIFT)–based image matching, Mask2Former segmentation, and skeleton‐based morphological analysis. SIFT identifies matched keypoints to retrieve relevant image pairs across inspections, Mask2Former with a Swin‐T backbone delivers superior segmentation accuracy, while a custom postprocessing pipeline quantifies crack width, length, and area using skeletonization and a branch‐pruning algorithm. These geometric features are compared over time to assess crack growth. Validation on ultra‐high‐performance concrete (UHPC) beam tests demonstrates robustness to variations in illumination, scale, and rotation, high segmentation performance (81.60% mIoU and 88.84% mAcc), and measurement error within 1 mm. The framework can support long‐term defect monitoring and enable consistent, data‐driven maintenance strategies in structural health monitoring.