With the development of UAV technology, autonomous aerial refueling technology increases the flight radius and payload of UAV and improves the combat effectiveness of UAV. The paper focuses on the precise guidance technology of UAV hose aerial refueling in complex environment, and studies the key link of drogue detection during the close docking phase of UAV autonomous aerial refueling. Using deep learning and graphics processing unit, a new method based on Faster R-CNN neural network is proposed. In order to ensure its robustness and wide application, an image deep learning data set was made by using real data of hose aerial refueling. Based on the experimental data, the robustness and identification accuracy of the identification algorithm based on Caffe framework Faster R-CNN cone sleeve were verified, and the comparison experiment proved that the identification algorithm also had better identification ability of cone sleeve in the complex UAV oil-feeding environment.