Vera Esteban, Guzmán Felipe, Weinberger Camilo
Appl Opt. 2021 Apr 1;60(10):B119-B124. doi: 10.1364/AO.417574.
The deep learning wavefront sensor (DLWFS) allows the direct estimate of Zernike coefficients of aberrated wavefronts from intensity images. The main drawback of this approach is related to the use of massive convolutional neural networks (CNNs) that are lengthy to train or estimate. In this paper, we explore several options to reduce both the training and estimation time. First, we develop a CNN that can be rapidly trained without compromising accuracy. Second, we explore the effects given smaller input image sizes and different amounts of Zernike modes to be estimated. Our simulation results demonstrate that the proposed network using images of either 8×8, 16×16, or 32×32 will dramatically reduce training time and even boost the estimation accuracy of Zernike coefficients. From our experimental results, we can confirm that a 16×16 DLWFS can be quickly trained and is able to estimate the first 12 Zernike coefficients-skipping piston, tip, and tilt-without sacrificing accuracy and significantly speeding up the prediction time to facilitate low-cost, real-time adaptive optics systems.
深度学习波前传感器（DLWFS）能够直接从强度图像中估计像差波前的泽尼克系数。这种方法的主要缺点与使用大规模卷积神经网络（CNN）有关，这类网络训练或估计耗时较长。在本文中，我们探索了几种减少训练和估计时间的方法。首先，我们开发了一种能够在不影响准确性的情况下快速训练的CNN。其次，我们研究了较小输入图像尺寸和不同数量待估计泽尼克模式所带来的影响。我们的模拟结果表明，所提出的使用8×8、16×16或32×32图像的网络将显著减少训练时间，甚至提高泽尼克系数的估计精度。从我们的实验结果可以证实，16×16的DLWFS能够快速训练，并且能够在不牺牲准确性的情况下估计前12个泽尼克系数（跳过活塞项、倾斜项和像散项），并显著加快预测时间，以促进低成本实时自适应光学系统的发展。
