Enhanced Receptive Field and Multi-Branch Feature Extraction in YOLO for Bridge Surface Defect Detection

Unmanned Aerial Vehicles (UAVs) are increasingly utilized for bridge inspections and play a crucial role in detecting defects. Nevertheless, accurately identifying defects at various scales in complex contexts remains a significant challenge. To address this issue, we propose RDS-YOLO, an advanced algorithm based on YOLOv8n, designed to enhance small-scale defect detection through the integration of shallow, high-resolution features. The introduction of the RFW (Receptive Field Weighting) module dynamically expands the receptive field and balances multi-scale detection accuracy. Additionally, the DSF-Bottneck (Dilated Separable Fusion) module further optimizes feature extraction, emphasizing the representation of small defects against complex backgrounds. The SA-Head (Shuffle Attentio) module, with shared parameters, precisely localizes defect zones while reducing computational costs. Furthermore, the EigenCAM technique improves the interpretability of the model’s output, offering valuable insights for maintenance and monitoring tasks. The experimental results demonstrate that RDS-YOLO outperforms YOLOv8n, achieving a 3.7% increase in average detection precision and a 6.7% improvement in small defect detection accuracy.