Benchmarking YOLOv8 Variants with Transfer Learning for Real-Time Detection and Classification of Road Cracks and Potholes

Road damage, including potholes and cracks, is a significant issue frequently encountered in road infrastructure in many regions. Such conditions accelerate road degradation, increase the risk of traffic accidents, and significantly increase the maintenance and repair costs. Although several deep learning models have been proposed for road damage detection, few studies have systematically compared the performance of lightweight YOLOv8 variants using a consistent dataset. To address this gap, this study proposes a road defect detection and classification model based on the YOLOv8 series, which is enhanced using transfer learning to improve performance and efficiency. The dataset, obtained from Roboflow, comprises 3,846 images categorized into training, validation, and testing sets. Three YOLOv8 variants—YOLOv8n, YOLOv8s, and YOLOv8m—were benchmarked for performance. A performance evaluation was performed using the metrics of precision, recall, and mean Average Precision (mAP). Results show that YOLOv8m achieved the highest precision (99.00%), recall (98.40%), and mAP (99.50%). In the pothole category, precision reached 98.70% and recall 99.30%; in the crack category, precision was 99.30% and recall 97.60%. The findings demonstrate that YOLOv8, particularly the YOLOv8m variant, is highly effective for real-time road damage detection and classification, offering a viable solution for intelligent transportation systems and automated infrastructure monitoring. This research has the potential to revolutionize infrastructure monitoring by enabling scalable, real-time, and cost-effective assessments of road conditions. It minimizes reliance on manual inspections, reduces human errors, and contributes to the development of intelligent transportation systems and predictive maintenance strategies.