A lightweight vision-based inspection method for wind turbine blade defects using UAV imagery

Abstract Visual inspection based on unmanned aerial vehicle (UAV) imagery has become a crucial non-destructive evaluation technique for wind turbine blades. However, reliable automated defect detection remains difficult due to extreme scale variations, low-contrast damage patterns, and cluttered optical backgrounds in aerial scenarios. Furthermore, practical deployment in UAV measurement systems requires highly compact algorithms for efficient edge-device inference. To address these challenges, this work proposes a lightweight vision-based inspection method, termed MSW-YOLO11n, which optimizes a baseline detection architecture with three complementary improvements for accurate defect evaluation. First, a multi-kernel grouped-convolution module (M_C3k2) is designed to enrich receptive-field diversity, effectively capturing the irregular morphological features of physical defects. Second, a parameter-efficient bidirectional feature pyramid (SBiFPN) is introduced to strengthen high-resolution pathways, ensuring robust multi-scale target fusion. Third, an outlier-aware bounding-box regression strategy (WIoU v3) with dynamic non-monotonic reweighting is applied to improve localization robustness under complex backgrounds. Experimental evaluations on a wind turbine blade defect dataset (dirt and damage) demonstrate that the proposed method achieves an optimized accuracy-compactness trade-off. MSW-YOLO11n improves the mAP@0.5 from 84.7% to 87.0% while reducing the number of parameters from 2.58 M to 2.26 M. Notably, the accuracy for the challenging structural damage category increases by 5.0% points, indicating improved detection reliability for edge-deployable UAV inspection tasks.