YOLOv8-PaveDet: Lightweight UAV Pavement Distress Detection With Context-Enhanced Features and Class-Aware Reweighting

High-resolution pavement imagery collected by uncrewed aerial vehicles (UAVs) enables scalable, non-intrusive road distress inspection. Combined with lightweight detectors, it supports flexible automated detection of surface distress across large road networks. However, accurate detection remains challenging because thin cracks are visually subtle and easily fragmented, scale variation across UAV altitudes is pronounced, and safety-critical categories such as potholes are severely underrepresented. Existing approaches address these challenges only in isolation: context-enhanced Transformer architectures achieve strong feature modeling but are computationally prohibitive for high-resolution UAV tiles; lightweight detectors preserve efficiency but sacrifice contextual aggregation; and existing long-tail loss functions are designed for generic benchmarks without accounting for the joint difficulty of thin-structure localization and severe class imbalance in UAV pavement scenes. To address these limitations, we propose YOLOv8-PaveDet, a task-oriented extension of YOLOv8n with three coordinated components: GhostConv for efficient backbone computation, ConvFormer-enhanced C2f blocks for long-range spatial aggregation over elongated crack structures, and a lightweight shared-channel detection head (LSCD) for scale-consistent multi-level prediction. A hybrid training loss combining IoU-aware SlideLoss with class-weighted binary cross-entropy provides explicit minority-class rebalancing. Experiments on UAV-PDD show that the architectural modifications raise [email protected] from 62.9% to 73.4% and [email protected]:0.95 from 32.4% to 42.5%, while reducing parameters from 3.01M to 1.83M and GFLOPs from 4.10 to 2.78. Adding the hybrid loss further improves [email protected] to 81.68% (Aggressive) and [email protected]:0.95 to 45.54% (Conservative), with Pothole [email protected] rising from 39.70% to 75.81%. These results demonstrate that coordinated lightweight design, combining contextual aggregation with explicit class rebalancing, achieves a stronger accuracy–efficiency balance for UAV pavement inspection; the resulting detections can serve as upstream sensing inputs for public infrastructure management.