LBSD-YOLO: A Lightweight YOLOv10-Based Network with Multi-Attention Enhancement for Bridge Surface Defect Detection

Bridge surface defect detection plays a critical role in ensuring traffic safety and facilitating infrastructure maintenance. A lightweight object detection network based on YOLOv10, termed LBSD-YOLO, is developed to achieve high detection accuracy while maintaining high efficiency for deployment on resource-constrained devices. The proposed framework consists of three main components: a feature extraction backbone, a feature fusion neck, and a detection head. In the backbone, the C2f\_FEMA (C2f with Feature Enhancement and Multi-branch Attention) module and the LAEDS (Lightweight Adaptive Encoder–Decoder for Sampling) spatial attention module are incorporated to enhance multi-scale feature representation. The neck incorporates multi-scale feature fusion with an Efficient Multi-scale Attention (EMA) mechanism. In the detection head, a lightweight DP-Head structure is developed, variant integrated with the DAMF\_CA coordinate attention for improved channel and spatial focus. Experiments are conducted on the self-built BDD-1234 dataset, which contains 6,617 high-resolution images covering six common bridge defect categories (cracks, spalling, exposed reinforcement, rust stains, efflorescence, and delamination). Compared to the baseline YOLOv10s, LBSD-YOLO reduces model size from 16.6 MB to 9.6 MB (42.2% reduction), computational complexity from 21.4 GFLOPs to 17.3 GFLOPs (19.2% reduction), and parameters from 7.2 M to 4.6 M (36.1% reduction), while achieving competitive detection performance with only a marginal decrease in overall mAP@50 (64.1% vs. 65.5%). The results demonstrate that LBSD-YOLO offers an efficient and accurate solution for real-time bridge defect detection on portable devices.