Instance Segmentation for Road Damage Identification Using YOLOv8

Road infrastructure plays a crucial role in ensuring safety, accessibility, and mobility, particularly for vulnerable groups such as the elderly and individuals who utilize wheelchairs. Nonetheless, deteriorating road conditions, such as potholes, cracks, and other forms of damage, present significant risks to road users. Traditional manual inspection methods are characterized by being time-consuming, labor-intensive, and frequently inconsistent, thereby necessitating the implementation of automated solutions for precise and efficient road damage detection. While computer vision-based approaches have demonstrated promising outcomes, challenges such as limitations in datasets, a variety of damage types, and the requirement for accurate pixel-level segmentation continue to persist without resolution. This study addresses this gap by proposing an instance segmentation framework using YOLOv8 to identify road defects with high granularity. To enhance the robustness of the model, we assess data augmentation techniques and dataset fusion strategies. Additionally, we conduct a performance comparison of YOLOv8 with YOLOv9 and YOLOv11 to ascertain the most effective model for this task. Experimental findings indicate that YOLOv8 achieves the highest average precision (mAP50) of 69.8, surpassing both YOLOv9 and YOLOv11 (each at 67.3). Significant enhancements through horizontal flip, 90° counter-rotation, and ±15° rotation are examined to evaluate their effect on detection accuracy. This research contributes to the progression of automated road defect assessment, offering insights into optimal model selection and augmentation strategies for practical implementation.