EMSA-YOLO: An Enhanced Multiscale Attention Model for Pavement Defect Detection

Abstract Accurate and efficient identification of pavement defects is crucial for reducing traffic accident rates and ensuring the safety of road users’ lives and property. To effectively address the multiscale issue in pavement defect detection and significantly improve the detection accuracy of pavement defects, this study takes the YOLOv8 model as the basic framework. A new detection model is designed through targeted improvements, which is named Enhanced Multiscale Attention-YOLO (EMSA-YOLO), which integrates multiscale attention mechanisms with large-target feature extraction capabilities. Firstly, to enhance the extraction of defect features across multiple spatial scales, the EMSA module is designed to expand the receptive field and incorporate multiscale contextual information through the application of a learned weight matrix. Secondly, DySample is employed to optimize the upsampling process in the neck network, yielding clearer fused feature images and enhancing the detection accuracy of small-sized potholes. Finally, we add a new P6 large target detection layer to expand the receptive field and capture multiscale features, which enables the model to capture deeper semantic features and improve the detection of longitudinal and transverse cracks. Experimental results on the RDD2022 dataset demonstrate that the EMSA-YOLO achieves 84.1% mAP@50 and 57.5% mAP@50-95, outperforming the baseline by 2.1% and 1.7%, respectively. Significant accuracy improvements are observed in three critical defect categories: transverse crack with a 2.4% increase, alligator crack with a 4.8% increase, and potholes with a 1.4% increase compared to the baseline performance.