Automated Crack Segmentation and Width Quantification Using Deep Learning for Structural Health Monitoring of Reinforced Concrete Structures

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Accurate characterization of crack geometry is critical for assessing the durability and serviceability of Reinforced Concrete (RC) structures. Although Deep Learning (DL) has significantly improved automated crack segmentation, its effectiveness in supporting engineering measurements such as crack width remains insufficiently explored. This study proposes an integrated image-based comparative study combining semantic segmentation with quantitative crack width estimation. Four deep learning models, U-Net, Attention U-Net, DeepLabV3+, and YOLOv8-seg were evaluated using a dataset of 2000 annotated crack images under consistent training conditions. Segmentation performance was assessed using standard pixel-level metrics, while crack width predictions were validated against measurements obtained from laboratory-tested RC beams. Results show that encoder–decoder architectures achieve superior boundary delineation, leading to more reliable width estimation. U-Net and Attention U-Net demonstrated the best overall performance with an Intersection over Union (IoU) of 0.741 and 0.740, respectively. In contrast, YOLOv8-seg exhibited significantly lower segmentation accuracy and higher measurement error. The findings confirm that segmentation boundary precision plays a critical role in extracting accurate structural parameters, highlighting the suitability of encoder–decoder models for vision-based Structural Health Monitoring (SHM) applications.