Real-time concrete crack segmentation for bridge structural health monitoring: A lightweight YOLOv11-based approach with multi-scale feature fusion

• DWR module fuses multi-scale cracks via dilated convs (1,3,5). ↑mAP50 to 0.930 • LAWDS cuts model 23% (42MB) via depth-wise conv + attention. Noise suppression • 62 FPS on edge (Jetson AGX). 44% faster than SOTA. Real-time deployment • >95% precision; detects 0.2mm cracks under rain/varying light. SOTA accuracy Existing concrete crack segmentation algorithms face critical bottlenecks in real-time bridge health monitoring, including insufficient accuracy for small/thin cracks, heavy computational burden, and poor robustness to complex on-site environments (e.g., shadows, stains). To address these issues, this study proposes an enhanced YOLOv11-based approach integrating two novel modules tailored for crack detection: the Dilation-wise Residual (DWR) module leverages parallel dilated convolutions to capture multi-scale crack features, specifically enhancing small/thin crack detection by fusing fine-grained details and contextual information, while the Light Adaptive-weight Downsampling (LAWDS) module combines depth-wise separable convolutions with an attention mechanism to suppress background noise and reduce model complexity. Ablation studies confirm the synergistic effect of these modules: compared to the baseline YOLOv11, the proposed model achieves a precision of 0.95 (+11.8%), recall of 0.93 (+5.9%), and mAP50 of 0.93 (+7.3%), with a 23% reduction in model size. Comparative experiments with state-of-the-art methods (Unireplknet, Fasternet, EfficientViT) further demonstrate its superiority in balancing accuracy and efficiency, while its 58 FPS inference speed on edge devices enables real-time on-site inspection. This work provides a novel technical solution for reconciling high accuracy and lightweight performance in bridge crack segmentation, laying a solid foundation for efficient structural health monitoring of bridges. –