Hgm-YOLOv10: A Heterogeneous Gradient Multi-Attention Approach for Wind Turbine Surface Defect Detection via drone

In the task of target detection of surface defects of wind turbines using drones, there are problems such as the lack of a suitable publicly available multi-category labeled dataset, multiple and densely distributed targets, and low drone arithmetic piggybacking. On this basis, a multi-category wind turbine surface defect detection dataset is constructed, and an algorithm Hgm-YOLOV10 (Heterogeneous Gradient Multi-Attention YOLOv10) is proposed to improve YOLOv10. A Hg-Dy (Heterogeneous Gradient Dynamic Attention Module) module is designed to replace one of the C2f modules. This enables the backbone network to capture a wider range of contextual information, extract multi-scale features, enhance feature representation, and improve detection performance. Secondly, a new detection head DSM (Dynamic Multi-scale head) is introduced to further improve the adaptability to complex scenes and increase the detection accuracy while maintaining the number of parameters stable. We also conduct experiments on the self-constructed dataset, and the original YOLOv10 algorithm improves 5.9% for map50 and 3.3% for map50-95.