HDS-YOLO: A high-performance real-time framework for wind turbine blade defect detection via synergistic feature preservation

The defect detection of wind turbine blades is a critical component for ensuring the structural integrity and operational safety of wind farms. Currently, this process increasingly relies on Unmanned Aerial Vehicle (UAV) inspection imagery; however, existing object detection algorithms still face severe challenges in handling the extreme scale variations between minute cracks and large-scale surface dirt, as well as in suppressing complex environmental background noise. To address these issues, this study has developed a high-precision, multi-scale defect detection model based on the YOLOv11 architecture, termed HDS-YOLO. The model introduces two synergistic improvement mechanisms to reinforce feature retention and spatial resolution. First, a Hybrid Down-sampling (HDS) module is integrated into the backbone network. Through a dual-stream design, it utilizes Space-to-Depth (SPD) operations to achieve lossless pixel rearrangement at the physical level, while incorporating a parallel convolutional branch as a low-pass filter to smooth high-frequency noise, thereby significantly improving the signal-to-noise ratio of input features. Second, a P2 high-resolution detection head is introduced to provide a fine spatial sampling rate that matches the output features of the HDS module, which substantially enhances the localization accuracy for millimeter-scale weak cracks. Experimental results demonstrated that the baseline YOLOv11 performed excellently in the Pareto balance of accuracy and efficiency, achieving an mAP of 80.80% and an inference speed of 526 F PS, which significantly outperformed Faster R-CNN and SSD. Furthermore, ablation experiments verified the effectiveness of HDS-YOLO. Compared to the baseline YOLOv11, the mAP@50 increased by 1.51%, Precision rose to 81.24%, and Recall improved by 4.14%. These results indicate that the proposed HDS-YOLO model provides a robust and efficient technical solution for the real-time intelligent inspection of wind turbine blades.