End-to-end subpixel positioning method for space dim and weak point targets based on multi-task learning

Abstract

Subpixel positioning of dim targets is critical for space optical navigation. To overcome the limited adaptability of threshold-based methods and error accumulation in staged deep learning approaches, we propose an end-to-end multi-task network. The model integrates Squeeze-and-Excitation (SE) and Hybrid Attention (HA) modules into a U-Net backbone, jointly generating a pixel-wise mask and a distribution map for centroid positioning. Subpixel coordinates are directly calculated from these outputs, eliminating error propagation between detection and positioning stages. On simulated data, our method achieves a 30% higher detection rate and 0.5 pixels lower RMSE compared to adaptive thresholding, and outperforms segmentation-only StarNet by 8.8% in positioning accuracy. Real-data experiments demonstrate a detection rate of 90.83% and an RMSE of 0.2826 pixels, confirming high robustness and subpixel precision.