Star sensors measure attitude by identifying stars in space missions,and star-map identification algorithms,as the core part of the sensors,determine the accuracy of the star sensors'attitude measurement.To ad-dress concerns that existing neural network-based star-map identification algorithms hardly reduce computational costs while guaranteeing identification accuracy,this paper proposes an end-to-end lightweight network star-map identifi-cation algorithm(i.e.,MobileCiT)based on an attention mechanism to directly identify noisy star-maps in star sen-sors.MobileCiT employs depthwise separable convolution and an improved pre-inverted residual structure based on a convolutional neural network.It also uses an attention mechanism to focus on the position information of star points.In addition,because of the high cost and uncontrollable noise of real star-maps,a coordinate mapping model based on small hole imaging is used to generate noisy simulated star-map training and test datasets.The experimental results show that the identification accuracy of MobileCiT for different noisy star-maps is 99.850%,which is higher than those of existing star-map identification algorithms based on the lightweight networks MobileNet and MobileViT.Moreover,it has good robustness to positional and magnitude noises,as well as false and missing stars.MobileCiT realizes high-accuracy star-map identification without the need for preprocessing operations,such as background denoising,connectivity domain detection,and star centroid extraction.MobileCiT improves the identification accu-racy with low computational costs,and the computation load is one-third of the algorithm based on the MobileViT network.MobileCiT is compared with star-map identification algorithms based on subgraph isomorphism or pattern recognition.Under the same field of view and noise conditions,MobileCiT showes higher identification accuracy and robustness,further verifying its superiority over traditional star-map identification algorithms.
关键词
星图识别/注意力机制/轻量化/星图仿真/卷积神经网络/噪声鲁棒性
Key words
star-map identification/attention mechanism/lightweight/star-map simulation/convolutional neural network(CNN)/robustness to noise
维普
万方数据