To address the problems of false detection and low detection accuracy of traditional wind turbine paddle detection algorithms in complex environments, a wind turbine paddle defect detection method integrating multi-scale features and attention mechanism is proposed. Firstly, the improved backbone network L-ResNet50 is used for feature extraction to retain more effective information. Then the attention mechanism module is embedded for different scale feature layers to enhance the focused semantic information. Finally, the extracted deep features and shallow features are fused with multi-scale features to further improve the model accuracy. Through the defect detection experiments on the wind turbine paddle images captured by UAV aerial photography, the results show that the average accuracy of the proposed method in the detection of wind turbine paddle defects in complex environments is improved by 8.2% compared with the original Faster R-CNN model.