Abstract: In the process of insulator defect inspection, it is difficult for traditional algorithms to take into account of detection accuracy and model size at the same time due to complex background. In this paper, an insulator defect detection model based on improved YOLOv5s is proposed. Firstly, a Bottleneck CSP structure and an attention mechanism of lightweight spatial and channel convolution are used to strengthen insulator characteristics and suppress the complex background characteristics. Secondly, an improved BiFPN structure is proposed to achieve multi-scale feature fusion and improve the ability of small object detection. Finally, K-means++ algorithm is used to re-cluster the prior frames, and lightweight GhostC3 and Ghost Conv modules are designed to ensure accuracy of the network and reduce size of the model. The experimental results show that the mAP of the improved algorithm in this paper reaches 92.3% in Insulator2022 dataset, with an increase of 3.6%; the number of parameters reduces by 26.73%, the floating point arithmetic reduces by 23.17%, and the missed detection rate reduces by 5.47%. In the open dataset, the defective insulator mAP reaches 99.5%. All of the evaluation index values are better than those of the mainstream algorithms of Faster-RCNN, SSD, YOLOv3 and YOLOv3-tiny as well as the related algorithms of insulator detection.