Abstract: With the development of smart grids, the aerial insulator defect detection based on computer vision is widely used in power inspection. In this paper, aiming at the low accuracy of the deep learning model for insulator self- explosion defect detection, and the difficult deployment to the mobile devices, such as drones, due to the large model, the YOLOv5s (You Only Look Once Version-5s) model is selected as the basic network to improve the detection accuracy, and the improved network is pruned to lighten the model. First, the SiLU activation function is replaced with the Mish activation function with a better gradient flow to enhance the network stability; second, the CBAM (Convolutional Block Attention Module) attention mechanism is fused into the last layer of the backbone feature extraction network to filter out more useful features; finally, the Transformer coding structure is embedded into the C3 module, and the C3 in the YOLOv5s feature fusion network is replaced with the new C3TR to strengthen the feature fusion capabilities of the high and low-level networks. The comprehensive pruning method is used for the improved model, and the redundant channels and convolution kernels in the network are cut out respectively to make the model more lightened. Through an experimental verification and compared with the current commonly used models on the test set, the detection accuracy of the improved model in this paper reaches up to 97.23%, the size of the model after pruning is only 0.5MB, the detection time is 1.8ms, and the number of floating-point operations is 0.61 G, which better meets the requirements of real-time detection of the transmission lines.