A multi-task deep learning network for robust and high accuracy star sensor image centroiding

Star sensors require reliable onboard centroid estimation despite binary-star interference and non-Gaussian distortions. Conventional algorithms lack robustness, while deep learning models often sacrifice accuracy or computational efficiency for real-time operation. This study proposes a lightweight multi-task CNN based on the HorNet baseline for onboard deployment. It simultaneously distinguishes single-star from binary-star targets and estimates sub-pixel centroids in a single forward pass. Two improvement plug-in modules—differentiable Gaussian convolution for noise suppression and coordinate attention for spatial enhancement—add minimal overhead, enhancing the centroid estimation accuracy for star spots of various morphologies. A Pareto-optimal loss function balances classification and regression without manual tuning. Experiments on a dataset (including ideal Gaussian, real non-Gaussian profiles, and binary-star interference) show that the proposed model achieves 100.0% binary-star detection rate, substantially outperforming conventional algorithms (≈ 53.7% accuracy). Under simulated non-Gaussian conditions, it attains a centroid error of ≈ 0.017 pixels, outperforming the conventional centroid method (≈ 0.16 pixels) and Gaussian fitting (≈ 0.12 pixels). The model remains compact, efficient, and easy to train, offering a structurally efficient solution with demonstrated potential for high-precision, real-time processing stellar centroid estimation on current GPU platforms, demonstrating potential for future onboard deployment.