远场光斑反演波前相位的深度学习方法

自适应光学系统中，波前传感器的准确性和鲁棒性极大地影响像差探测能力和闭环校正效果。在波前振幅分布不均匀或信标光能量不足的情况下，哈特曼波前传感器由于存在子孔径缺光现象会导致传感精度下降，而基于远场光斑反演波前相位的无波前传感自适应系统实时性难以满足实用需求。基于深度学习复原波前的方法是通过输入远场光强图像直接求取像差，可以作为自适应光学系统的有效补充。文中通过数值模拟，证明了深度残差神经网络能够通过远场光斑直接预测波前相位的Zernike系数。实验验证了输入与重构波前相位之间校正后残差RMS为0.08λ，GPU加速后的平均计算耗时小于2 ms。该方法能较准确地预测入射波前畸变的Zernike系数，具有一定像差校正能力，适合在传统自适应光学技术中，用于测量并校正波前畸变的主要成分，或为优化式自适应光学提供良好的初始波前估计。

Method of inverting wavefront phase from far-field spot based on deep learning

In the adaptive optics system, the accuracy and robustness of wavefront sensor greatly affect the ability of aberration detection and closed-loop correction. In the condition of the nonuniformity of amplitude distributions or insufficient of beacon light energy, it will cause accuracy decrease of Hartmann wavefront sensing due to the lack of sub-aperture light. Meanwhile, the real-time performance of the wavefront sensing-free adaptive system based on far-field spot inversion cannot meet the practical requirements. The method of the wavefront inversion based on deep learning is to directly obtain aberrations by inputting the far-field light intensity image, which can be used as an effective supplement to the adaptive optical system. Through numerical simulation, this paper proved that the deep residual neural network could directly predict the Zernike coefficient of the wavefront phase through the far-field spot. And experimental demonstrated the corrected residual RMS between input and reconstructed wavefront phase was 0.08 waves, the average computation time was less than 2 ms by GPU acceleration. This method can predict the Zernike coefficient of incident wavefront distortion more accurately, and has a good aberration correction capability, suitable for measuring and correcting the main components of wavefront distortion in traditional adaptive optics method, or providing a good initial wavefront estimation for optimized adaptive optics.