<![CDATA[Road surface damage, including cracks and structural cavities, poses significant risks to vehicle safety and sustainable infrastructure management. Conventional road assessment remains largely manual, time-consuming, and subjective, creating a need for automated and scalable solutions. This study presents a deep learning framework for pixel-level pavement damage segmentation using an in-house dataset of 3,000 annotated road images collected under diverse real-world conditions. The framework comprises image preprocessing, model training, and standardized performance evaluation. Three established segmentation models, namely U-Net, DeepLabV3+, and YOLOv8-seg, were comparatively evaluated under consistent experimental settings using Intersection over Union (IoU), Dice coefficient, and inference speed. The results show that all three models can identify damaged road regions, although their performance varies according to architectural characteristics and damage patterns. U-Net achieved the highest IoU and Dice scores, demonstrating strong capability in preserving spatial information and delineating fine-grained damage, particularly thin cracks. DeepLabV3+ produced competitive results in segmenting larger and context-dependent damage areas through multi-scale feature extraction. YOLOv8-seg provided faster inference and greater potential for real-time applications, although its segmentation accuracy was lower than that of U-Net and DeepLabV3+. These findings demonstrate that model selection should consider the trade-off between segmentation accuracy and computational efficiency. U-Net was the most effective model for precise pixel-level road damage segmentation in this dataset. Future research will focus on broader datasets, additional architectures, and deployment-oriented optimization for intelligent road maintenance systems.]]>
