【摘要】 星敏感器工作在复杂的空间环境，强噪声干扰将严重影响其姿态测量的效果。星像提取是星敏感器星图识别和姿态估算的必要前提，研究抗干扰的快速星像提取算法是提高星敏感器性能的有效途径。结合星敏感器星像目标特点，提出基于改进Faster R-CNN的星敏感器抗干扰快速星像提取算法。首先，在研究Faster R-CNN的基础上，通过构建星像特征提取网络，优化FPN和RPN结构，实现星像快速粗提取，确定各星像所在区域。然后，提出基于像素筛选的星像质心精提取算法，计算高精度的星像质心坐标，最终实现强噪声干扰环境下的快速星像提取。利用星敏感器仿真方法建立星图数据集，开展以星像特征提取网络为主干网的星像提取网络训练和星敏感器星像提取实验。结果表明，在添加概率分布分别为50和0.08的泊松-高斯复合噪声条件下，提出算法的星像目标识别率达到97.6%,对于1 024×1 024像元的单幅星图，平均处理时间小于30 ms,星像提取精度达到0.03个像元，优于扫描法和矢量法。更多还原

【Abstract】 The star sensors work in a complex space environment. The strong noise interference will affect the effectiveness of their attitude measurement. Star image extraction is a necessary prerequisite for star pattern recognition and attitude estimation of star sensors. Studying anti-interference and fast star image extraction algorithms is an effective way to improve the performance of star sensors. An anti-interference and fast star image extraction algorithm for star sensors based on the improved Faster R-CNN is proposed. Firstly, on the basis of studying Faster R-CNN,by establishing a star image feature extraction network and optimizing the RPN and the FPN,star images are extracted coarsely and rapidly. Star image regions can be determined. Then, a pixel-based and fine star image centroid extraction algorithm is presented. High-precision star image centroid coordinates are calculated. The fast star image extraction is realized at last in the case of the strong noise interference. A star map dataset is established by using the star sensor simulation method. The star image extraction network with the star image feature extraction network as the backbone is trained. The star image extraction experiments are carried out. It shows that when the Poisson-Gaussian mixed noises with the probability distributions of 50(Poisson)and 0.08(Gaussian)respectively are added, the proposed algorithm achieves a star image target recognition rate of 97.6%. For a single star map with 1 024×1 024 pixels, the average processing time is less than 30 ms.The star image extraction accuracy reaches 0.03 pixels. The algorithm is better than the scanning method and the vector method.更多还原