The recognition and attitude estimation of non-cooperative spacecraft targets in geostationary orbit (GEO) is the premise of on-orbit service actions. However, it is difficult to identify the on-orbit state of the non-cooperative spacecraft. The prior information for the target is lacking and the spacecraft is maneuvering and tumbling during the action. In this paper, the deep learning method based on machine vision is used to detect and recognize the key components of the non-cooperative spacecraft. The image sample library of the spacecraft is established, and the key components in the image are marked. After the training in the faster region-based convolutional neural network, the name and position of the key components on the image can be obtained by using the method. The mean average precision (MAP) reaches 82%. The work in this paper is significant for relative attitude measurement, on-orbit service operation, multi-source fusion perception, and so on.