Phase diversity algorithm with high noise robust based on deep denoising convolutional neural network.

The wave-front phase expanded on the Zernike polynomials is estimated from a pair of images by the use of a maximum-likelihood approach, the in-focus image and the defocus image, which contaminated by noise, will greatly reduce the solution accuracy of the phase diversity (PD) algorithm. In the study, we introduce the deep denoising convolutional neural networks (DnCNNs) into the image preprocessing of PD to denoise the in-focus image and defocus the image containing gaussian white noise to improve the robustness of PD to noise. The simulation results show that the composite PD algorithm with DnCNNs is better than the traditional PD algorithm in both RMSE of phase estimation and SSIM, and the mean of the RMSE of the phase estimation of the improved PD algorithm is reduced by 78.48%, 82.35%, 71.09% and 73.67% compared with the mean of the RMSE of the phase estimation of the traditional PD algorithm. The well-trained DnCNNs runs fast, which does not increase the running time of traditional PD algorithms, and the compound approach may be widely used in various domains, such as the measurements of intrinsic aberrations in optical systems and compensations for atmospheric turbulence.