AMSW-YOLO: A Lightweight Multimodule and Loss Function Synergistic Optimization Algorithm for Intelligent Pavement Defect Detection

Abstract Pavement defect detection is essential for ensuring road structural integrity and traffic safety. Traditional visual inspections and sensor-based methods suffer from efficiency and cost limitations, while existing deep-learning-based approaches still face practical constraints. This study proposes an improved AMSW-you only look once (YOLO) object detection algorithm, constructs the Multisource Road Defect Dataset (MRDD) for systematic evaluation, and further assesses cross-scene generalization on the RDD2022 data set. Through a synergistic optimization strategy integrating multiple modules and loss functions, the model significantly enhances feature representation while maintaining a lightweight design [2.7M parameters, 7.4G floating-point operations (FLOPs), 5.6MB model size]. On MRDD, AMSW-YOLO improves mean average precision (mAP)50, mAP50-95, and F 1 -score from 0.936, 0.612, and 0.910 to 0.957, 0.642, and 0.940, with crack AP increasing by 4.4%. On RDD2022, mAP50, mAP50-95, and F 1 -score increase by 5.7%, 5.0%, and 7.5%, respectively, with pothole AP showing the most significant improvement of 8.3%.