Yang Sen, Li Xiaofeng
Opt Express. 2022 Oct 10;30(21):37874-37887. doi: 10.1364/OE.470595.
Deep neural networks have contributed to the progress of image-based wavefront sensing adaptive optics (AO) with the non-iterative regression of aberration. However, algorithms relying on the one-shot point spread function (PSF) typically yield less accuracy. Thus, this paper proposes an iterative closed-loop framework for wavefront aberration estimation outperforming the non-iterative baseline methods with the same computation. Specifically, we simulate the defocus PSF concerning the estimation of the Zernike coefficients and input it into the backbone network with the ground-truth defocus PSF. The difference between the ground-truth and estimated Zernike coefficients is used as a new label for training the model. The prediction updates the estimation, and the accuracy refined through iterations. The experimental results demonstrate that the iterative framework improves the accuracy of the existing networks. Furthermore, we challenge our scheme with the multi-shot phase diversity method trained with baseline networks, highlighting that the framework improves the one-shot accuracy to the multi-shot level without noise.
深度神经网络通过像差的非迭代回归推动了基于图像的波前传感自适应光学（AO）的发展。然而，依赖一次性点扩散函数（PSF）的算法通常精度较低。因此，本文提出了一种迭代闭环框架，用于波前像差估计，在相同计算量下性能优于非迭代基线方法。具体而言，我们针对泽尼克系数估计模拟离焦PSF，并将其与真实离焦PSF一起输入主干网络。真实泽尼克系数与估计的泽尼克系数之间的差异用作训练模型的新标签。预测更新估计，通过迭代提高精度。实验结果表明，迭代框架提高了现有网络的精度。此外，我们用基于基线网络训练的多次曝光相位多样性方法对我们的方案进行了挑战，突出表明该框架在无噪声情况下将一次性精度提高到了多次曝光水平。
