Abstract: Aiming at the interference of shooting angle changes and insignificant features in the drone aerial photography environment,this paper proposed an improved GoogLeNet convolutional neural network to identify and locate the wind turbines,which can automatically extract wind turbine category features without manual pre-selection.The deep feature vectors of wind turbines are constructed through GoogLeNet network.In the network model training process,the concept of transfer learning is introduced and the pre-trained GoogLeNet network is trained by using wind turbine images.The classification network can be prevented from falling into the local optimal solution while speeding up the model training.The region proposal network and the multi-task loss function are used to integrate the candidate region search and border regression into the network in the Faster RCNN framework,so that the wind turbines in the aerial image can be automatically classified and annotated,and the time complexity can be reduced.Experimental results show that the optimized GoogLeNet network can improve the accuracy of target visual detection in the complex aerial photography environment and complete the task of wind turbine automatic positioning by means of transfer learning.The avera-ge accuracy of wind turbines based on GoogLeNet is over 96%.

Key words: Convolutional neural network, Deep learning, GoogLeNet model, Transfer learning, Visual inspection, Wind turbine