Artificial neural network for star tracker centroid computation

Abstract

We propose a unique dataset with star images, their centroids, and a new centroid algorithm based on machine learning, that significantly improves star image centroid performance. The centroid of the star image is the subpixel coordinates on the image plane corresponding to the assumed point source of the light (a star). We have compiled and open-sourced an unparalleled dataset of 50000 images of bright stars with their sub-pixel coordinates. It can be used by scientists to refine various techniques and algorithms related to photogrammetry, astrometry, and photometry. Data for this dataset are obtained from the Gaia space observatory, the most recent generation of space astrometric missions. The proposed algorithm outperforms the classical Center of gravity method by almost an order of magnitude.