<p>A wind turbine blade surface defect detection model based on YOLOv5s is proposed to overcome the limitations of existing models in exploiting surface defect characteristics. It focuses on precise multi-scale defect target perception in complex backgrounds. The Convolution for Wind Turbine Blades (Conv_WTB) module is designed to capture multi-scale features, ensuring information integration and transmission, and addressing the issues of traditional receptive fields and alignment errors. The deformable spatially adaptive attention module is developed to enhance key feature sensitivity by utilizing deformable convolution and global average pooling to improve detection in complex scenes. The Smoothed Intersection over Union (SIoU) is introduced to optimize boundary errors and enhance defect target localization. Experiments conducted on a four-class defect dataset demonstrate that the improved model boosts F1-score by 4.04% to 96.58% and mAP@0.5 by 3.94% to 98.10%. It outperforms other existing models. When tested on the NVIDIA GeForce RTX 3090 GPU, the FPS reaches 41.67 frames·s⁻¹, meeting real-time requirements. This model provides support for the intelligent operation and maintenance of wind power generation units.</p>