Implementasi Deep Learning untuk Sistem Deteksi Jalan Berlubang dengan Komparasi Metode YOLOv8 dan YOLOv9

Abstract

This research implements and comparatively evaluates deep learning algorithms YOLOv8 and YOLOv9 for an automated classification system detecting road pothole damage using digital images. Road damage, with its complex visual characteristics, presents significant image processing challenges. The dataset comprises four damage classes: large deep, large shallow, small deep, and small shallow, collected via mobile devices and expanded through data augmentation (rotation, flipping, lighting adjustments). The development process included acquisition, preprocessing, model training with PyTorch (100 epochs, batch size 16), and performance evaluation using mAP-50, precision, recall, and F1-score. The study results indicate that YOLOv9 marginally outperformed in accuracy, achieving a mAP-50 of 0.92, precision of 0.98, recall of 0.98, and an F1-score of 0.97. YOLOv8, however, demonstrated very good performance with a mAP-50 of 0.91, precision of 0.96, recall of 0.97, and an F1-score of 0.96. Despite YOLOv9's higher accuracy, it required substantially longer training time (2.4 hours) compared to YOLOv8 (0.49 hours) and faced batch size limitations due to GPU memory constraints. Primary misclassifications occurred between visually similar classes. This system is expected to provide a fast, efficient, and automated solution for road infrastructure monitoring, enhancing user safety, and reducing maintenance costs. Further development potential includes model fusion or ensemble techniques for optimizing accuracy and resource efficiency.