Technique for Recovering Wavefront Phase Bad Points by Deep Learning

Abstract

In adaptive optics systems, the bad spot detected by the wavefront detector affects the wavefront reconstruction accuracy. A convolutional neural network (CNN) model is established to estimate the missing information on bad points, reduce the reconstruction error of the distorted wavefront. By training 10,000 groups of spot array images and the corresponding 30th order Zernike coefficient samples, learns the relationship between the light intensity image and the Zernike coefficient, and predicts the Zernike mode coefficient based on the spot array image to restore the wavefront. Following the wavefront restoration of 1,000 groups of test set samples, the root mean square (RMS) error between the predicted value and the real value was maintained at approximately 0.2 μ m. Field wavefront correction experiments were carried out on three links of 600 m, 1.3 km, and 10 km. The wavefront peak-to-valley values corrected by the CNN decreased from 12.964 μ m, 13.958 μm, and 31.310 μm to 0.425 μm, 3.061 μm, and 11.156 μm, respectively, and the RMS values decreased from 2.156 μm, 9.158 μm, and 12.949 μm to approximately 0.166 μm, 0.852 μm, and 6.963 μm, respectively. The results show that the CNN method predicts the missing wavefront information of the sub-aperture from the bad spot image, reduces the wavefront restoration error, and improves the wavefront correction performance.