Comparative Performance and Edge Deployment of Quantized YOLO Architectures for Wind Turbine Defect Detection

Unmanned aerial vehicles (UAVs) combined with artificial intelligence can automate the inspection of wind turbine blades. This benchmark and deployment study compares lightweight YOLO detectors (YOLOv8, YOLO11, YOLO26; nano and small sizes; AdamW and SGD optimizers) trained for single-class damage detection on 5763 blade images and deployed on a Raspberry Pi 4. Each configuration was trained with three seeds and evaluated on a test split (mean ± standard deviation). At 640 × 640 pixels the six SGD configurations reached 75.5–76.8% mAP@50 after quantization and cannot be ranked by accuracy; with the default hyperparameters SGD gave a higher recall than AdamW in all 18 seed-matched pairs. Mixed-precision post-training quantization (backbone and neck INT8, detection head FP16) lowered mAP@50 by 0.8 percentage points on average and made the model files about three times smaller, but on the Raspberry Pi 4 it increased the latency by 1.5–2.2 times compared with FP32. YOLO26n-SGD, the fastest nano model, reached, in FP32, 9.9 FPS, but only 70.5 ± 0.5% mAP@50 at 320 × 320 (model trained at that size), against 2.6 FPS and 77.0 ± 0.9% at 640 × 640. SSDLite320 and EfficientDet-Lite0 reached 25.2% and 40.1% mAP@50. Flight tests and power consumption were not examined.