Deep learning for surface crack detection in civil engineering: A comprehensive review

Abstract

Ensuring the safety of civil engineering structures requires the adoption of advanced crack detection techniques, with deep learning offering a highly effective solution to overcome the limitations of traditional manual inspection methods. This systematic review critically evaluates contemporary deep learning techniques for detecting surface cracks in civil structures, highlighting their applications and challenges. Beginning with an analysis of publicly available crack datasets and evaluation metrics, the study lays a foundation for advancing crack detection research. It then examines various deep learning approaches, including classification, recognition, and segmentation tasks. Special focus is given to applications in key civil engineering domains such as roads, bridges, and tunnels. Notably, the DCUFormer model achieved the highest average Intersection over Union (mIoU) of 0.821 and an F1-Score of 0.894 on the Crack500 dataset, positioning it as the most effective model for road structure crack detection to date. The review also explores persistent challenges, including dataset scarcity and imbalance, evaluation method standardization, accurate damage assessment, crack propagation prediction, and the integration of real-time, multi-source sensing systems with edge AI technologies. By providing a comprehensive overview of the state-of-the-art in deep learning for crack detection, this paper enhances the understanding of its capabilities and limitations, offering valuable insights for scholars and industry practitioners in this critical research area.