As a critical component of wind turbines, defects in wind turbine blades pose a significant threat to their operation. To improve the precision and recall rates of defect detection in wind turbine blades, this paper proposes an Efficient Multi-Scale Convolution module (EMSConv) to replace the convolutional modules in residual blocks for grouped convolution, targeting the YOLOv8n network. Multiple attention mechanisms from Dynamic Head are introduced in the detection head, leveraging the synergy between multiple self-attention mechanisms across feature layers for scale, spatial, and task awareness, thereby enhancing the representational capability of the object detection module. By integrating Inner-IoU, Wise-IoU, and MPDIoU, a novel Inner-Wise-MPDIoU is proposed to replace CIoU, improving the network"s detection accuracy. Tested on a custom dataset of wind turbine blade defects, the experiments show that the proposed YOLOv8-EDI achieves a mAP50 value of 81.0% on this dataset, a 2.3% improvement over the original YOLOv8n; the recall rate reached 76.8%, a 3.7% improvement; and the computational complexity of the network structure was reduced by 5.5%. Compared with the original model, YOLOv8-EDI exhibits superior localization ability and detection accuracy for wind turbine blade defects, meeting the demands of industrial-scale batch detection.