Automatic crack detection and segmentation of masonry structure based on deep learning network and edge detection

Abstract

Because of high time expense, strong experience dependence and labor intensity, traditional crack detection methods have given way to computer vision-based automatic non-destructive methods, which have been successfully applied to the crack detection of concrete structures. However, the presence of mortar joints between masonry blocks introduces considerable noise in masonry images compared to concrete images, resulting in less efficiency and accuracy of the computer vision-based methods in masonry structures. This study aims to incorporate deep learning network into computer vision to achieve rapid and accurate crack detection in masonry structures. Firstly, the GELAN-Seg network, i.e. the YOLOv9-Seg network without the Auxiliary Supervisory Branch, was improved by adding an edge processing branch to obtain the YOLO-Edge model for higher accuracy of crack segmentation. Then, low-cycle reversed loading experiments were conducted on masonry walls to collect crack images, which were further manually labelled to create a comprehensive dataset of cracks in masonry structures. To assess the effectiveness of the proposed model, the crack segmentation performance of the YOLO-Edge model was compared with GELAN-Seg, YOLOv9-Seg, Mask R-CNN, SOLOv2, YOLACT, ConvNeXt, and Swin Transformer. The results demonstrate that the proposed model can significantly improve crack segmentation accuracy with AP0.5_seg reaching 83.8 %, outperforming other models, e.g. GELAN-Seg (80.4 %), YOLOv9-Seg (81.6 %), Mask R-CNN (64.5 %), ConvNeXt (70.0 %), YOLACT (67.9 %), Swin Transformer (68.5 %) and SOLOv2 (80.7 %). Additionally, the proposed model has smaller parameters, and thus less computation and higher detection speed than other models. These findings confirm the efficacy of the YOLO-Edge model in the crack detection for masonry structures.