Comparative Analysis of Deep Learning Models for Road Damage Segmentation

다국어 초록 (Multilingual Abstract)

Road pavement deterioration directly affects traffic safety and long-term infrastructure sustainability, necessitating accurate detection and quantitative assessment of damaged regions. While bounding-box-based object detection can localize defects such as cracks and potholes, it is limited in representing irregular boundaries and estimating precise damaged areas. This study conducts a controlled comparative evaluation of road-damage instance segmentation models under identical experimental conditions. Specifically, two onestage architectures (YOLOv8n-seg and YOLOv11-seg) and a representative two-stage framework (Mask RCNN) are analyzed using a dataset of 4,567 real-world pavement images annotated with pixel-level masks.
Performance is assessed in terms of Precision, Recall, F1-score, mAP50, mAP50–95, and model parameter scale. Experimental results show that YOLOv11-seg achieves the highest precision, indicating effective suppression of false positives, while YOLOv8n-seg maintains comparable overall accuracy with a lightweight structure. Mask R-CNN demonstrates high recall but requires substantially larger parameter capacity. The findings highlight how architectural differences influence precision–recall balance, boundary stability, and computational efficiency in quantitative pavement damage analysis.