Single far field type deep learning wavefront restoration method based on four-quadrant discrete phase modulation

Abstract

The invention discloses a single far field type deep learning wavefront restoration method based on four-quadrant discrete phase modulation. Deep learning can perform self-extraction of image deep features, has strong nonlinear fitting capability, and can be used for the fitting mapping of far-field light intensity distribution to wavefront aberration information. In Fourier optics, the far-fieldlight intensity distribution is equal to modulus square of incident wavefront complex amplitude Fourier transform, and the single far-field light intensity distribution can correspond to a plurality of different incident wavefronts. In supervised learning, equivalently, one sample corresponds to multiple tags, and the mapping is ill-conditioned state mapping. A focal plane wavefront recovery sensor which is simple in structure, high in light energy utilization rate, good in real-time performance and high in recovery precision is designed, four-quadrant discrete phase modulation is introduced,and the multi-solution problem that a single far field corresponds to multiple incident wavefronts in wavefront recovery is solved; and the mapping relationship between the modulated far-field light intensity distribution and the incident wavefront is accurately fitted by using deep learning, so that high-precision rapid wavefront restoration of a single-frame focal plane light intensity image isrealized.

Description

technical field [0001] The invention relates to a wavefront restoration method, in particular to a single far-field deep learning wavefront restoration method based on four-quadrant discrete phase modulation. Background technique [0002] In Fourier optics, the far-field light intensity distribution is equal to the square of the modulus of the complex amplitude Fourier transform of the incident wavefront. At this time, a single far-field light intensity distribution can correspond to multiple different incident wavefronts. In supervised learning, this is equivalent to one sample corresponding to multiple labels, which is an extremely pathological mapping. As a branch of supervised learning, deep learning cannot learn this kind of pathological many-to-one mapping. Under small aberrations, the difference between multiple different incident wavefronts corresponding to a single far-field light intensity distribution is very small, so deep learning can still be used to achieve hi...

AI Technical Summary

Problems solved by technology

Kong Qingfeng, Institute of Optoelectronics Technology, Chinese Academy of Sciences, analyzed the causes of the multi-solution problem and used the iterative optimization method to restore the wavefront, see [Kong Qingfeng. Research on wavefront phase inversion method based on single-frame focal plane image[D]. University of Electronic Science and Technology of China, 2019], compared with the traditional phase inversion method, this method reduces the number of algorithm iterations, but the real-time improvement of the algorithm is limited

Images