Research on lightweight detection of road defects based on improved YOLOv10n model

In recent years, deep learning techniques based on convolutional neural networks have advanced rapidly, with their image processing capabilities widely applied in object detection and semantic segmentation. Addressing the challenges of small object misdetection and background interference in automated road surface defect detection, this paper proposes an improved lightweight YOLOv10n model integrating C2fMamba and CBAM. C2fMamba models long-range dependencies within the detector head at linear complexity, overcoming the limitations of convolutional receptive fields. CBAM dual attention is concatenated to adaptively focus on key channels and spatial regions, suppressing redundant features. A proprietary RDD2022ES dataset comprising 25,598 images covering eight typical defect categories was constructed. After training at 640×640 resolution with Mosaic augmentation and 100 epochs of SGD with cosine annealing, the improved model achieved mAP50 of 0.461 (up from 0.438) and mAP50-95 of 0.261 (up from 0.246), with only a marginal increase in parameters. The confusion matrix reveals diagonal confidence scores exceeding 0.53 for both longitudinal and transverse cracks. Recall rates for potholes and subsidence cracks improved by 6.3 and 6.8 percentage points respectively, with false alarm rates below 0.04. The PR curve maintains high recall at 0.91 while significantly suppressing background misclassifications. Visualisation results indicate that the average confidence score improved by 0.05–0.1, with detection box edges better aligning with actual damage. Experiments confirm this method achieves a new equilibrium between accuracy, robustness, and efficiency, enabling direct deployment on drones or vehicle-mounted edge devices. It provides a practical lightweight AI solution for intelligent road maintenance.