Numerical Research On Machine Learning Assisted Direct Wavefront Sensing Method

Over the years,optical microscopy has become an essential tool in biomedical research.Fluorescent proteins have been used to label tissues,and the fluorescent signals excited by the laser are collected for imaging.Due to the inhomogeneous distribution of refractive index in biological tissues,the laser beam is strongly scattered during the propagation and leading to the low imaging quality.Adaptive optics is a kind of wavefront compensating method which can recover diffraction-limited performance for optical microscopes.Adaptive optics can be classified into direct and indirect wavefront sensing methods.Direct wavefront sensing method based on Shack-Hartmann wavefront sensors is widely used in in vivo imaging because of its advantages in time consumption.But the traditional algorithms cannot reconstruct complex wavefronts accurately.In this article,we propose two direct wavefront sensing methods combining with machine learning for complex wavefronts reconstruction.The wavefronts are characterized by Zernike polynomials,and then we established a mapping model between the distorted Shack-Hartmann patterns and the corresponding Zernike coefficients.Compared with the traditional wavefront reconstruction algorithm in direct wavefront sensing methods,the detection accuracy and the speed are improved by the proposed method.Since Zernike polynomials have limitations in characterization of complex wavefronts,especially for high-frequency details,we propose a new method for complex wavefronts characterization.And we divided distorted Shack-Hartmann patterns into small regions and utilized two convolutional neural networks to reconstruct local wavefronts and then stitching the whole wavefronts.Compared with several kinds of direct,indirect and machine learning assisted wavefront sensing methods,our algorithm has significant strength in complex wavefronts detection and a better generalization ability.