基于卷积神经网络的高精度分块镜共相检测方法

High-precision co-phase method for segments based on a convolutional neural network

为获得与单口径望远镜相当的空间分辨率,使成像系统成像质量达到或接近衍射极限,拼接主镜式望远镜的分块子镜应确保实现共相位拼接,本文针对拼接主镜式望远镜高精度平移(piston)误差检测问题,提出了一种基于卷积神经网络的高精度平移误差检测方法.通过在成像系统的出瞳面上设置具有离散孔的光阑,构建了对平移误差极为敏感的点扩散函数图像数据集,根据此数据集的特点搭建了具有高性能的网络模型,并测试得到网络的最佳检测范围.仿真结果表明,在略小于一个波长的捕获范围内,单个网络能够准确地输出一个或多个分块子镜的平移误差;应用于六子镜成像系统时,平移误差检测精度达0.0013λRMS(root mean square),并且方法对残余倾斜(tip-tilt)误差、波前像差、CCD噪声、光源带宽具有良好的鲁棒性.该方法简单快速,可广泛应用于分块镜系统的平移误差检测.

In order to achieve the resolution comparable to the resolution of a monolithic primary mirror telescope and make the imaging quality of the imaging system reach or approach to the diffraction limit,the submirrors of the segments telescope should ensure co-phase splicing.To solve the problem of phase error detection,a highprecision piston error detection method is proposed based on convolutional neural network(CNN).By setting a mask with a sparse multi-subpupil configuration on the exit pupil of the imaging system,a point spread function(PSF)image dataset that is extremely sensitive to the piston error is constructed.According to the characteristics of this dataset,a high-performance CNN model is built.And the best detection range of CNN is tested.The simulation results show that a single network can accurately output the piston error of one or more submirrors in the capture range slightly less than one wavelength.When the single network is applied to the six-submirror imaging system,the detection precision of the piston error reaches an RMS value of 0.0013λ(here,RMS stands for root mean square).And the method has good robustness to residual tip-tilt error,wavefront aberration,and CCD noise,light source bandwidth.The method is simple and fast,and can be widely used to detect the piston error of the segments.