Early and effective detection of wind turbine blade(WTB) surface defects is crucial for enhancing operational efficiency and ensuring the safety of wind power generation systems. Deep learning technology has made significant progress in WTB defect detection in recent years, particularly in improving detection precision and achieving real-time monitoring. However, there are still the following challenges in dealing with various small target complex surface defect detection scenarios of WTB: (1) the existing models have limited selective attention to input features, especially when dealing with complex backgrounds and small target defects of WTB. (2) the ability to handle low-quality images in WTB detection is insufficient. This paper proposes a high-precision model LSC-YOLO for small target defect detection based on YOLOv9. Firstly, LSC-YOLO employs a Large Selective Kernel attention (LSK) that dynamically adjusts the receptive field, more effectively addressing the differences in background information required for the detection of small targets. Then, SPD-Conv and CARAF instead of the original downsampling and upsampling module, respectively, to alleviate the second problem. The LSC-YOLO algorithm shows improvements across multiple metrics on the wind turbine blade dataset, with a multi-threshold average precision ( \(mAP\_0.5:0.95\) ) reaching 74.33%, which is 3.8% higher than YOLOv9, and an increase of 0.03 in \(F1-confidence\) . To investigate the model’s versatility and generalization, we included additional datasets for further study, and the results demonstrate that our enhanced algorithm performs optimally, achieving \(mAP\_0.5:0.95\) of 70.73%, surpassing YOLOv9 by 4.5%. This provides a new solution for the precision and reliability of small target detection and offers valuable insights for future related research.