Abstract: The centroid measurement accuracy and computational efficiency of star spots in star charts are key performance indicators for star sensors in space. This study aims to develop a deep learning-based centroid measurement method (Deep Learning-based Centroid Measurement, DLCM) to address the limitations of traditional centroid measurement methods in terms of accuracy and computational efficiency, particularly under noisy conditions and complex star charts. The DLCM method utilizes convolutional neural networks (CNN) to automatically extract complex features from star charts, and employs multiple fully connected layers in the output layer of the network to predict the centroid position through regression. To train the neural network, Gaussian spots under various noise levels are simulated, and the network structure is optimized using a large volume of training data. The DLCM method adapts to varying noise levels and image conditions without requiring manual parameter adjustments or preprocessing based on image characteristics. Experimental results demonstrate that the DLCM method achieves a centroid measurement accuracy of 0.05 pixels within a 3σ range, with excellent robustness and generalization capabilities. Furthermore, DLCM shows significant advantages in computational efficiency. The experimental results validate the potential application of DLCM in star chart centroid measurement, showcasing its high accuracy and efficiency. This method provides effective technical support for the development of future high-precision star sensors and other electro-optical pointing measurement devices.

Key words: deep learning, star images, centroid measurement algorithm, convolutional neural network, star sensors