Research on LD-YOLO for Surface Defect Detection of Wind Turbine Blades

To address the challenges of large target scale differences and complex background interference in surface defect detection of wind turbine blades, this paper proposes a lightweight improved YOLOv8 detection framework, named LD-YOLO. The main work include: (1) Using the Ghost lightweight module to compress redundant features, significantly reducing computational load while ensuring accuracy; (2) Designing a BiFPN-GLSA fusion structure, combining bidirectional multi-scale feature pyramid and global-local spatial attention mechanism, to enhance the ability to locate defects under background interference; (3) Proposing the SPPF-LSKA module, which realizes intelligent expansion of the receptive field through large-kernel selective attention convolution; (4) Utilizing the SEAMHead detection head, optimizing feature expression through spatial-channel collaborative attention. Experiments on the Wind-turbine-blade-surface-defect dataset demonstrate that the optimized model reduces the number of parameters by 49.5%, decreases computational load by 40.2%, and achieves a 0.8% improvement in detection accuracy compared to the original YOLOv8. This framework provides an efficient, lightweight, and practical solution for real-time defect detection of wind turbine blades.