<![CDATA[Automated road damage detection using computer vision technology remains an active research challenge, particularly do to issues of imbalanced datasets and variations in road surface conditions. While deep learning (YOLO) methods have been widely applied, a comprehensive performance comparison of the latest architectures and the impact of data balancing strategies require further analysis. This study aims to compare the performance of the latest YOLO generations (YOLOv8, YOLOv9, YOLOv10, and YOLO11) and conduct an ablation study to determine the influence of on-the-fly oversampling. The research began by collecting road damage image data (pothole, crack, and manhole) from the public source: "potholes, cracks and openmanholes (Road Hazards)" on Kaggle. Following the data cleaning process, experiments were conducted using 1845 images. The dataset split ratio for train:valid:test sets was 70:15:15. The dataset was then trained using the four YOLO architectures across two case studies: an imbalanced dataset and a balanced dataset achieved through oversampling on-the-fly of the minority classes. Model performance was evaluated using Precision, Recall, F1-Score, mAP50, and mAP50-95 metrics. The results show no significant indications of overfitting and confirm the models' strong generalization capabilities. The application of on-the-fly oversampling quantitatively improved the performance of all models, with the highest mAP50-95 increase observed in YOLO11m (+0.055). This strategy successfully improved the detection of minority classes (crack and manhole) without degrading performance on the majority class (pothole). In the balanced dataset scenario, YOLO11m (mAP50-95: 0.468) and YOLOv8m (mAP50-95: 0.467) demonstrated the best overall performance. A unique finding was shown by YOLOv9m, which achieved the highest Precision value of 0.914.]]>
