For the problems of small objects of insulator defects and insignificant texture features in UAV power inspection images, a multi-stage insulator defect detection method is proposed based on MSR-Net and YOLOv5-s. Firstly, the YOLOv5-s network is applied to quickly identify the insulators in the images. Secondly, a multi-scale super-resolution network (MSR-Net) is proposed to reconstruct the insulators image. Due to the introduction of residual structure, normalized coordinate grid, and multi-scale data mixed training strategy, the high-frequency feature extraction capability and multi-scale generalization capability of the network have been improved, which can significantly enhance the characteristics of insulator defect small objects. Finally, input high-resolution image into YOLOv5-s network again to reidentify insulator defect more effectively. The test results on the self-built datasets show that the average detection accuracy and average detection speed of the proposed insulator defect detection method reach 88.04% and 5.24FPS, respectively, which effectively improves the accuracy of small object detection of insulator defects. The proposed method can be widely used in power inspection and other small object detection scenarios.