Morphology-aware supervision improves real-time wind turbine blade defect localization in unstructured natural environments

Unmanned aerial vehicle (UAV) inspection is widely used for wind turbine blade maintenance, yet detecting minor damage in operational wind farms remains difficult because background clutter, extreme defect aspect ratios, and curved or bifurcated image-space geometries all reduce accuracy. We propose SWCB-YOLO, a lightweight edge detector built on YOLOv11n. Its central contribution is a Curvature-Aware Shape IoU (CA-Shape-IoU) loss, a morphology-aware image-space localization regularizer whose supervision derives from crack morphology: Bézier-curvature, skeleton-distance, and elongation fields are precomputed offline from polygon masks and applied during training, leaving the detection head and inference cost unchanged. The loss is paired with an Asymmetric Strip Swin-TransformerV2 (AS-Swin2) module for long-range dependencies along slender cracks and a Frequency-Spatial Dual-Domain Attention (FS-DDA) module that combines DCT-informed channel recalibration with spatial attention. On a field dataset spanning a full annual cycle, SWCB-YOLO attains 89.0% mAP@50 and 52.8% mAP@50:95, improving on the YOLOv11n baseline by 13.8 and 11.7 percentage points under a common training protocol, with 3.70M parameters at 35 FPS on a Jetson Xavier NX. The improvement remains positive in every turbine-grouped cross-validation fold (+12.4 to +14.6 points) and after near-duplicate test frames are removed. Evaluation on five open benchmarks, including two blade-defect datasets, further examines transfer beyond the development site; the implementation and benchmark configurations are publicly available.