YOLOv8 for Road DamageRecognition: Deep Learning-Based Segmentation and Detection System

Road surface detection is critical for improving traffic safety and reducing road maintenance costs. Because traditional methods are time-consuming and costly, deep learning-based image processing techniques offer an important alternative in this field. This study aims to develop a model that automates the detection and segmentation of road surface defects such as potholes, manhole covers, and culverts using deep-learning-based image processing techniques. In this study, a dataset previously used in the literature was preferred. It was observed that object detection was performed using a dataset from the literature. In this study, both object detection and object segmentation were performed using different parameters. To prove the success of object segmentation, both object detection and segmentation were performed using the YOLOv8 algorithm, which has previously obtained successful results. AdamW optimization and Auto Batch parameters were selected for this study. With these parameters, object detection was first performed with the YOLOv8s model, which is one of the variances of the YOLOv8 algorithm with the most successful results in the literature, and a successful 92.8% mAP@50 performance value was obtained according to the sources in the literature. In this study, the YOLOv8n, YOLOv8s, YOLOv8m, and YOLOv8l variance models of the YOLOv8 algorithm were used with the preferred parameters, and segmentation was performed. In object segmentation, a map@50 performance value of 90.9% in all classes and 99.1% in culverts was obtained using the YOLOv8l model. A map@50 performance value of 89.1% for pothole segmentation and 88% for manhole cover segmentation was obtained using the YOLOv8s model. The results of the analyses showed consistency in precision and recall values. These findings contribute significantly to improving road safety, reducing maintenance costs, and supporting sustainable urban infrastructure. Future research could explore integrating multiple data sources and adapt these models to more complex road conditions.