An Integrated Framework for Segmentation and Management of DIC Cracks in Concrete Structures Using Synthetic Data, Unsupervised Learning, and BIM

Digital image correlation (DIC) is a crucial technique for capturing and monitoring cracks in concrete during laboratory structural tests. However, existing deep learning-based crack detection methods predominantly rely on supervised learning, which is costly due to the need to collect DIC datasets through destructive structural tests. Furthermore, conventional 2D-based DIC image analysis is inadequate for dynamic visualization and analysis of crack developments in 3D structures. To address these challenges, this study introduces an integrated framework for automated segmentation and management of DIC-based cracks using synthetic data, unsupervised learning, and building information modeling (BIM). First, a deep learning-based synthetic image generation method is presented to simulate realistic DIC cracks without structural tests. Second, an unsupervised learning network, enhanced by a customized multi-stage data augmentation strategy, is proposed for label-free DIC crack segmentation. Third, a DIC-enriched BIM management system is developed to automatically integrate multimodal DIC data with 3D physical models of structures while enabling interactive data management. The performance of the proposed framework was assessed through both synthetic datasets and real-world three-point bending tests on three concrete beams. The results demonstrate that the proposed segmentation method outperforms existing SOTAs, and the developed system enables multimodal DIC data visualization for subsequent SHM applications.