Faster-RCNN Based Star Image Extraction Algorithm Suitable for Star Sensors Disturbed by Gaussian Noise

Star sensors are one of the most prominent attitude measurement devices nowadays. They can provide high-accuracy attitude information to space vehicles by observing stars in the sky to obtain star maps, extracting star images, and recognizing stars through star pattern identification and attitude estimation. Their excellent performance of strong independence and anti-electromagnetic interference has attracted interest all over the world. However, strong noise and stray lights in the severe working environment pose a challenge to star sensors. Faint star images and degraded star maps increase the difficulty of star image extraction, leading to potential risks of attitude measurement failure. In this paper, a novel star image extraction algorithm based on Faster-RCNN was proposed to improve the ability of star sensors to detect star images accurately when they were disturbed by noise. Firstly, the structure of Faster-RCNN was introduced, and ResNet-50 was selected as the backbone after comparing the features of three convolutional networks, i.e., MobilenetV2, Resnet50, and VGG16. The procedure of star image extraction was described in detail. Secondly, a star imaging simulation method was proposed and applied to establish the training and testing datasets. Then, experiments were conducted when Gaussian noise was considered during star imaging. The star extraction results, such as operation time and centroid accuracies of star images, were analyzed. It was found that the proposed algorithm has strong immunity to Gaussian noise interference. For star maps with $1024 \times 1024$ pixels, the average update rate is about 10 Hz.