为了提高复杂动态环境下精密稳像系统的质心定位精度,提出了一种基于生成对抗网络(GAN)的方法,以解决载体运动引起的星斑拖尾问题。传统的高斯模型先验方法对噪声敏感,盲反卷积方法采用简化模型估计模糊核,与真实值误差较大。所提方法结合了运动流图的运动方向和GAN的优化细节,将运动流图与模拟星点作为监督信号,采用端到端的形式复原星斑图,无需先验信息与迭代计算,并且可以抑制噪声。实验结果表明,相较于R-L方法,所提方法复原星斑后峰值信噪比提升了30.5%,质心定位精度提高了33.4%。

To improve the centroid localization accuracy of fine image stabilization systems in complex dynamic environments, this study proposes a method based on generative adversarial network (GAN) to solve the problem of star smearing caused by the satellite's carrier motion. The conventional Gaussian model method with prior information is sensitive to noise, and the blind deconvolution method estimates the blur kernel using a simplified model, leading to a large error compared with the real one. The proposed method combines motion direction of motion flow graphs and optimized details of GAN. The motion flow graphs and simulated star points are used as monitoring signals. The proposed method uses an end-to-end manner to restore star images without prior information and iterative calculations, and noise is suppressed. Experimental results show that the peak signal-to-noise ratio of star recovery and centroid localization accuracy are improved by 30.5% and 33.4%, respectively, compared with the R-L method.