Wavefront sensing using deep learning for Shack Hartmann and pyramidal sensors

The use of Deep Learning in wavefront sensing has become a tremendous tool that provides an innovative approach to estimate the phase of an aberrated wavefront. Different methods have been developed in this field in order to find the best strategy according to the application. In this paper, a comparison between two wavefront sensing applications is carried out using various CNN Deep Learning architectures to obtain the closest representation of it in terms of Zernike Coefficients. A database of 110 000 40x40 images is created per application using the OOMAO Matlab toolbox where the sensor's images were generated using 200 Zernike Coefficients to produce a wavefront. The proposed network SH-P Net for the Shack Hartmann and Pyramid sensor fits with lower RMSE than the others evaluated in this document. The database generated with uniform distribution allows the neural network to learn better and faster. The computation time of the networks was about 1 s for 22 000 images and the root mean square wavefront error between the estimation and the input was about 0.00664 rad approximately.