Aiming at the problem that the soft aerial refueling long-distance visual recognition cone sleeve is easy to lose details, we propose a kind of high-resolution detection layer based on the existing convolutional neural network of YOLOv8, which is used to fuse deep features and shallow features. We introduce the context aggregation and feature refinement network to synthesize the target context information and capture the feature information of the cone sleeve, which improves the detection probability of the key points of the long-distance cone sleeve target. The experimental results show that the proposed method is more convergent in key point difference loss and key point confidence loss compared with the original network, and has higher accuracy in target recognition and key point detection of small pixel area cone sleeve, and has 4.17% improvement in key point recognition accuracy, which meets the requirements of aerial refueling visual navigation.