CBAM-enhanced YOLOv11 for pavement crack detection: performance evaluation and ablation analysis

Abstract

Vehicle-mounted road images contain many crack-like structures, including lane markings, shadows, repaired asphalt strips, and low-contrast texture changes. These visual factors make crack localization difficult when the target is thin or partly discontinuous. This study evaluates a YOLOv11 detector for pavement crack recognition and examines whether a convolutional block attention module (CBAM) can improve feature selection in this setting. The experimental dataset consists of 2266 manually annotated inspection images and is expanded to 11,330 samples through rotation, cropping, Gaussian noise, color perturbation, and brightness adjustment. Three defect categories are used: longitudinal crack, transverse crack, and repaired crack. Two attention placement is compared with the original YOLOv11: CBAM inserted between the final C3k2 block and SPPF, and CBAM inserted between SPPF and C2PSA. All models are trained for 100 epochs with the same input size, batch size, optimizer, and hardware. The baseline reaches the highest mAP50 of 66.9%. The two CBAM variants obtain lower mAP50 values, but their F1 scores increase from 0.286 to 0.305 and 0.322, respectively. The result indicates that the tested attention module improves the precision–recall balance before it improves average precision. The current evidence therefore supports CBAM@SPPF as a promising placement for further training and tuning, rather than as a fully optimized replacement for the baseline.