Abstract: Aiming at the problems of large hardware resource consumption and slow recognition speed of image recognition algorithm, based on YOLOv5 algorithm, an insulator target detection model dedicated to power inspection UAV is designed. The convolution operation module and residual module in the algorithm are improved, and the learning depth of the algorithm is deepened by increasing the number of convolution layers. In order to improve the training speed, the learning and training of the data set is realized by using the multiple recurrent neural network training method. The fastest single image recognition speed of the model is 0.061 s, and the highest recognition accuracy of insulators is 98.9%.The results show that under the premise of consuming less hardware computing resources, the model can directly process the images collected by aerial photography, realize rapid identification, and can meet the requirements of real-time image processing in the process of power UAV inspection.