CBAM-ECA-YOLO: a deep learning framework for crack detection in civil engineering structures

Abstract

Crack detection in concrete structures is vital for ensuring structural integrity and public safety. Manual inspectionis labor-intensive and subjective, while existing deep learning methods often struggle to identify fine cracks under complex textures and varying illumination. To address these limitations, this paper presents WCE-YOLOv8, an enhanced attention-driven framework for accurate and efficient crack detection. The model introduces a grayscale channel alongside RGB input to strengthen texture contrast and emphasize subtle crack patterns. In the backbone, a Convolutional Block Attention Module (CBAM) refines spatial and channel information, while an Efficient Channel Attention (ECA) mechanism in the detection head improves lightweight feature representation. Additionally, a combined Wise-IoU and normalized wasserstein distance loss enhances bounding box regression for small and thin cracks. Experiments on the SDNET2018 dataset demonstrate that WCE-YOLOv8 achieves superior performance with precision of 85.34%, recall of 92.63%, F1-score of 87.20%, mAP50 of 89.28%, and mAP50–95 of 53.72%, outperforming YOLOv3, YOLOv5s, YOLOv7, and YOLOv8s. The proposed model delivers robust, real-time, and reliable crack detection, offering an effective non-destructive solution for monitoring and maintaining concrete infrastructures.