infraCrackNet: Automated Crack Detection in Infrastructure Using U-Net with Dataset Preparation, Manual Labeling, and Segmentation Performance

Early detection of structural failure provides safe-guards and protection to buildings and infrastructures. However, crack inspection on structures is labor-intensive and time-consuming. To address this issue, an annotated dataset is proposed to perform concrete surface crack detection using deep learning techniques. The dataset comprises 4,174 images, both with and without cracks, that were hand-labeled at pixel-level accuracy with segmentation masks to segment each area into cracked and non-cracked. We show the effectiveness of our dataset by training a U-Net model that achieved an accuracy of 98%. Both the trained model (with weights) and the labeled dataset are made publicly available for enhancing further research and practical applications of the system, such as automated surveillance and maintenance processes. This open release has the potential to contribute to accelerating progress in concrete surface crack detection toward more efficient and effective structural health monitoring systems.