基于卷积神经网络的自适应光学校正关键理论和方法研究

Adaptive optics (AO) is a powerful means to compensate wavefront distortion induced by atmospheric turbulence, system aberration or other factors. In the wavefront-sensorless AO system, the optimal corrector driving signals are obtained by iteration algorithm based on the measured light intensity. The system has the advantages of simple structure and low cost and is becoming a new trend in the development of AO extension and application. However, a certain number of iterative computation is needed in the correction process, which severely limits the real-time performance of the system. In this project, an AO correction method for non-iteratively restoring phase information is proposed through extracting features from intensity image using convolution neural network (CNN). The Zernike coefficients of wavefront or the driving signals of the corrector can be directly predicted by measuring the intensity image and the real-time performance of AO system can be effectively improved in this method. Research on key theory and method of wavefront-sensorless AO correction based on CNN will be implemented in this project. The theoretical basis and method of constructing CNN structure suitable for restoring phase from light intensity image will be explored. The feasibility of the AO correction method based on CNN will be verified by experiments. Finally, the correction ability of the system under the condition of laser atmospheric transmission is simulated, analyzed and evaluated. Based on the frontier of science and technology, the new method of image recognition in the field of artificial intelligence is applied to the problem of restoring phase from light intensity image in the field of AO in this research, which can provide new ideas for the development of new AO technology.

自适应光学(AO)是补偿由大气湍流、系统像差或其他因素造成波前畸变的有力手段。无波前传感器AO系统基于测量光强通过迭代算法寻找最优的校正器驱动信号，具有结构简单、成本低廉等优点，成为AO技术推广应用发展的新趋势，但校正过程需要大量迭代严重限制系统的实时性。本项目提出利用卷积神经网络(CNN)提取光强图像特征非迭代复原相位信息的AO校正方法,即测量光强图像直接预测波前Zernike系数或校正器驱动信号，可以有效提高AO系统的校正实时性。项目拟开展基于CNN的无波前传感器AO校正的关键理论和方法研究，重点探索基于CNN的光强图像复原相位的理论依据和方法，搭建基于CNN的AO校正实验，验证方法的可行性，最后仿真分析和评估在激光大气传输条件下该系统的校正能力。项目的研究立足科技前沿，将人工智能领域图像识别新方法应用于AO领域光强图像复原相位问题中，能够为新型AO技术发展提供新思路。

自适应光学(AO)是补偿由大气湍流、系统像差或其他因素造成波前畸变的有力手段。本项目提出利用卷积神经网络(CNN)提取光强图像特征非迭代复原相位信息的AO校正方法,即测量光强图像直接预测波前Zernike系数或校正器驱动信号，可以有效提高AO系统的校正实时性。.（1）项目构建和优化了适用于光强图像复原相位的CNN结构的理论和方法，研究表明使用Alexnet和GoogleNet等神经网络可以有效的实现畸变像差的校正；基于GoogleNet的研究可以发现，数据集大小和模型深度对训练结果有重要影响，当数据集较小时，较浅的神经网络会有较好的校正效果，随着数据集的增大，深层的神经网络校正效果有明显提高；噪声对模型预测结果有重要影响，当信噪比较低时模型预测效果较差。.（2）搭建了一套基于CNN的静态畸变校正AO实验系统，对比了Resnet50、Inceptionv1、Efficiententv2_m、Mobilenet_v3_large、Efficiententv2_m（迁移学习）及Mobilenet_v3_large（迁移学习）6种CNN结构的波前复原模型，研究表明基于CNN的AO校正系统可以实现仅靠两帧光强的畸变波前校正；网络的结构对波前校正的影响复杂，精巧的网络结构和适当的网络深度是实现畸变波前高精度校正的重要途径；使用迁移学习可以有效的改善模型预测精度，提高模型的泛化能力。.（3）建立了激光大气传输场景下基于CNN的AO校正模型，研究了系统在不同湍流条件下的校正规律。研究表明，湍流大气参数对基于CNN模型的校正结果有重要影响，湍流大气相干长度较小时，基于CNN的AO系统校正效果较好，随着湍流大气相干长度的增加校正效果逐渐变差；湍流像差中不同像差类型对基于CNN的AO系统的校正有明显影响，通过比较Zernike像差的远场强度模式，发现光强越集中，光斑图案越简单，误差越小；相反，光斑严重断裂，图案越复杂，误差越大。.项目的研究将人工智能领域图像识别新方法应用于AO领域光强图像复原相位问题中，能够为新型AO技术发展提供新思路，研究结果能够为实际激光传输工程中智能波前传感及新型自适应光学应用设计提供参考依据。