In the current deep space asteroid exploration missions, optical navigation is one of the main navigation systems, and autonomous target recognition is its highly critical part. Currently, motion analysis and light-curve analysis are two of the most widely used methods for autonomous target asteroid recognition. However, both of them are difficult to accurately identify the target within a short time (20 min). The traditional light-curve analysis method relies on the periodic brightness characteristics of asteroid. But in short-term observations, its brightness periodicity is not obvious, leading to the ineffectiveness of traditional methods, such as FFT and Lomb-Scargle, in distinguishing between stars and small celestial bodies.This paper focuses on the problem of autonomous target recognition in the autonomous optical navigation of the far-distant approach orbit phase in asteroid exploration missions and proposes a light-curve classifier based on Transformer, which can accurately distinguish between stars and small celestial body target from image of navigation camera within a short time (20 min).The Transformer-based classifier designed in this paper can extract both local and global features of the light-curve simultaneously. Its network structure retains only the encoder part of Transformer and adds linear layers and fully connected layers for the classification task. The encoder dynamically adjusts the weights through the multi-head self-attention mechanism to capture the periodic terms (local features) and trend terms (global features) in the input light-curves, thereby achieving the classification of stars and small celestial bodies.To train and test the classifier, this paper constructs simulation datasets and semi-simulation datasets. The simulation datasets are generated through computer graphics and photoelectric-link-simulation tools, simulating the light-curves of stallers and asteroid by star catalogue or target shape. The semi-simulation data integrates the DAMIT dataset and the MMT-9 dataset. Firstly, segmentation, fitting, and upsampling were performed on these light-curves to increase the sampling frequency, and short-time-interval and high-frequency simulation data were constructed using the photoelectric-link-simulation tool. The training strategy of the classifier adopts a pre-training and fine-tuning method: Firstly, pre-training is conducted using the simulation data, and then the classification layer is fine-tuned using the semi-simulation data.The classification results on the test dataset show that the proposed Transformer classifier achieves an average classification accuracy of over 95% on the simulation dataset, with a classification accuracy of 98.6% for stars and 95.4% for asteroid. On the in-orbit observation data of the OSIRIS-REx mission, compared with the Lomb-Scargle light-curve analysis and the motion analysis method, the Transformer classifier shows significant advantages. For the motion method, within a 20-min observation period, the positional changes of stars and target Bennu in the image are less than 5 pixels, making it difficult to distinguish and unable to complete the target recognition. Compared with the Lomb-Scargle algorithm, the Transformer classifier does not rely on the frequency domain features of the light-curve and can accurately distinguish between stars and non-star targets when lomb-scargle algorithm fail. Within a 20-min observation period, for the classification of target Bennu and background stars, the accuracy rates of the Transformer classifier are 92% and 83% respectively, much higher than 42% and 67% of the Lomb-Scargle algorithm.Based on the comprehensive experimental results, the proposed Transformer-based light-curve classifier exhibits excellent performance on both simulation and actual in-orbit data, effectively solving the problem of autonomous target recognition in the far-distant approaching orbit of small celestial body exploration missions within a short time. It significantly improves the accuracy of target recognition and provides an efficient technical approach for autonomous optical navigation in small celestial body exploration missions, possessing certain practical value and application potential.