It is very important for military defense and space exploration to automatically perceive the category and working state of space targets from the observation images. In order to realize the automatic and accurate perception of spatial target image information, a data-driven spatial target image information perception technology is proposed. The proposed technology is based on a deep convolutional neural network, using massive simulation data and a small amount of real data to train the neural network. The trained neural network can directly perceive information such as load and working status of targets from a spatial target image. Taking two spatial target image information perception tasks as examples, the technical practicability is tested. In the task of space target load recognition, the proposed technology can perform load recognition on unknown space target images with different degrees of blur and different noise levels. The results show that for different space target loads, the average recognition accuracy of the proposed technology exceeds 80%, and the detection speed can reach 50 frame/s. In the task of spatial target state perception, the model combination method is used to build the expert system of spatial target working state perception. Taking the space object image as an example, this paper realizes the perception of the working state information of the space object, and verifies the effectiveness of the data-driven space object image information sensing technology.