To achieve precise centroid localization in dynamic star images subjected to proton irradiation， a deep learning-based algorithm is proposed. A keypoint model for the centroids of streaked star spots is first established， followed by the construction of a star centroid keypoint detection network based on this model. The energy degradation behavior of dynamic star images and the effects of proton irradiation are considered， and a star image dataset is generated by integrating proton cumulative irradiation and transient experiments. This dataset， supplemented with actual irradiated dynamic star images， is employed to train and validate the keypoint detection network. The localization accuracy of the network is subsequently compared with conventional restoration-based methods.Validation on simulated non-irradiated star images demonstrates that， within an angular velocity range of 0 （°）/s to 7 （°）/s， the centroid localization error ranges between 0.03 and 0.2 pixel， surpassing restoration-based methods by approximately an order of magnitude. For actual irradiated dynamic star images with angular velocities between 2.5 （°）/s and 10 （°）/s， the localization error varies from 0.4 to 1.8 pixel. Furthermore， the proposed method effectively addresses challenging scenarios where traditional methods falter， such as distinguishing star spots from transient bright spots and resolving overlapping star streaks. The algorithm exhibits robust performance against radiation noise and high-dynamic interference， offering a novel solution for the application of star sensors in complex environments.