Due to the interference of the harsh space environment on satellites in orbit, the images outputed by the star sensor cameras in orbit are distorted compared to those calibrated on the ground. This distortion restricts the accuracy of attitude measurement, and real-time effective image correction is difficult due to the limitations of in-orbit operating conditions, thus limiting the accuracy of the star sensors. To address these issues, this paper proposes a method for on-orbit correction of star sensor distortions. The method involves constructing a mapping dataset between the distorted coordinates of star points and the theoretical coordinates from the star catalog to train a neural network model, which fits the nonlinear distortions in star sensor imaging, achieving high-precision image distortion correction. To verify the distortion correction capability of this method, a simulated star map experiment was conducted. Through the simulation experiment, the average error between the measured coordinates of the star points and the theoretical coordinates was reduced from 0.7237 pixels to 0.0586 pixels after correction. This result indicates that the proposed correction method can significantly enhance the resolution accuracy of star sensors and has the potential for on-orbit learning capabilities and extending to other onboard payload image correction applications.