Prediction of wavefront distortion for wavefront sensorless adaptive optics based on deep learning

Abstract

Aimed at the slow detection speed and low measurement accuracy of wavefront aberration in current wavefront sensorless adaptive optic technology, different convolution neural networks (CNNs) are established to detect the turbulence wavefront, including an ordinary convolutional neural network, a ResNet network, and an EfficientNet-B0 network. By using the nonlinear fitting ability of deep neural networks, the mapping relationship between Zernike coefficients and focal degraded image can be established. The simulation results show that the optimal network model after training can quickly and efficiently predict the Zernike coefficients directly from a single focal degraded image. The root-mean-square errors of the wavefront detection accuracy of the three networks are 0.075 λ , 0.058 λ , and 0.013 λ , and the time consumed for predicting the wavefront from the single degraded image are 2.3, 4.6, and 3.4 ms, respectively. Among the three networks presented, the EfficientNet-B0 CNN has obvious advantages in wavefront detection accuracy and speed under different turbulence intensities than the ordinary CNN and ResNet networks. Compared with the traditional method, the deep learning method has the advantages of high precision and fast speed, without iteration and the local minimum problem, when solving wavefront aberration.