Star recognition based on mixed star pattern and multilayer SOM neural network

This paper proposes a novel star recognition method based on a mixed star pattern and multilayer self-organizing map (SOM) neural network. The mixed star pattern (leverages known data, rather than having to be trained on raw images) consists of k stars' normalized relative angular distance with the invariance of proportion and rotation and the information of the stars' magnitude comparison, which can better describe the star pattern, increase the recognition rate, and reduce the average field of view (FOV). By adopting a SOM neural network-based hierarchical treelike structure, the proposed approach can improve the speed of star tracking, reduce the storage occupation for saving connection weight matrix of a neural network, and exploit the robustness of a SOM neural network when classifying a small amount of samples. The experimental results show that the proposed method takes up approximately 20% of the processing time of the traditional method (4489 stars) and performs better on both robustness and instantaneity than a modified grid algorithm based on a neural network under additive noise.