LMD_YOLO: A Lightweight and Efficient Model for Pavement Defects Detection

Abstract

Pavement defects detection is vital for ensuring safety and extending infrastructure lifespan. Traditional manual inspection methods are inefficient and insufficient for the increasing demand for road maintenance. While automated detection systems are essential, existing approaches face challenges such as complex road environments, diverse defect types, and high computational requirements for real-time processing. These methods often sacrifice either accuracy or efficiency. To address these challenges, a novel model, LMD_YOLO, is proposed. The model incorporates several innovations: the Diverse Branch Block enhances the detection head, improving accuracy; the mg_conv module replaces conventional convolution layers in the neck network, optimizing feature fusion without increasing computational cost; and improvements to the backbone network further enhance efficiency. These advancements enable LMD_YOLO to achieve high detection accuracy and robustness in complex conditions while maintaining computational efficiency. Experimental results show that LMD_YOLO outperforms existing models, with a 3.7% increase in mAP50 and 3.2% in mAP50-95. In terms of speed, LMD_YOLO achieves 63 frames per second, fully meeting real-time detection requirements. Additionally, its computational cost (FLOPs) is only 46.3% of YOLOv8n, demonstrating superior performance while significantly reducing resource consumption.