Deep learning self-image update procedure in a wavefront sensor based on the Talbot phenomenon under Gaussian illumination

Abstract

A feature-based image update procedure using machine learning is proposed to use in preprocessing of self-images in a Talbot wavefront sensor. A variant of the recurrent neural network with backpropagation, which is one of most widely applied machine learning tools, is utilized to stabilize intensity distribution in self-images in the case of an optical beam with a Gaussian profile. Once well trained, the neural network can decrease pit image shifts caused by beam intensity distribution in the case of a cosine-like grating. It is shown that based on the proposed recurrent neural network, it is possible to decrease the shift error caused by the Gaussian beam up to nine times depending on the aberration order and value. Despite the decreasing shift error, the value of the error of the restored aberration coefficient does not decrease significantly because of the feature-vector extraction method. It is shown additionally that due to the spatial spectrum wideness, the proposed self-image procedure is not applicable to binary gratings on the example of binary gratings with square pits. Adequate simulations are implemented to demonstrate the effectiveness and accuracy of the proposed approach.