Answers are grounded in this paper and cite their sources.
Pavement distress detection is critical for preventive maintenance. Unmanned aerial vehicles (UAVs) offer an efficient alternative for rapid pavement inspection. However, pixel-level segmentation faces challenges due to complex backgrounds, low contrast, large-scale variations, and high annotation costs. This study proposes a framework encompassing data generation and segmentation network design for UAV pavement scenes. To reduce reliance on pixel-level annotations, a label generator was developed to produce multiclass structural labels, which were then used to guide a fine-tuned ControlNet in synthesizing spatially aligned image–mask pairs for training. For segmentation, LiteSCSF-UNet is designed with shallow structural enhancement, long-range contextual modeling, and cross-level selective fusion. Experiments on the self-built UAV-Road dataset show that the model achieves 83.43% mIoU at 46.55 FPS with only 4.48M parameters; three-fold synthetic augmentation further improves mIoU to 87.92%. These results demonstrate that the proposed framework provides an effective solution for UAV-based pavement-scene segmentation under limited annotations.
No takes yet. Share an insight, caveat, or question.
Yan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: