Study On The Wavefront Sensing Technology In Adaptive Optics Based On Deep Learning

Adaptive optics system can correct wavefront distortion caused by atmospheric turbulence and image blur caused by inhomogeneous refractive index in various biological tissues,so it is widely used in astronomical observation,free space optical communication,biomedical imaging and other fields.As an important part of adaptive optics system,wavefront sensor provides phase information of distorted wavefront for wavefront control and correction in adaptive optics system.Therefore,it not only determines the correction accuracy of the system,but also affects the stability of the system to a great extent.Among them,Shack-Hartmann wavefront sensor and wavefront sensor based on interference are the two most commonly used wavefront sensors in adaptive optics system.They are also widely used in optical metrology and laser beam quality diagnosis,and have very important research value.At the same time,deep learning technology with artificial neural network as the core has developed rapidly in recent years,and has become one of the most successful and potential technologies in the field of artificial intelligence.Just as the so-called "artificial intelligence is the new electricity",deep learning technology,like electricity in the second industrial revolution,is rapidly popularized and applied in various scientific and industrial fields.The combination of deep learning technology and adaptive optics wavefront detection,wavefront control and other aspects is also being widely and deeply studied,which has broad development potential.This dissertation focuses on the use of deep learning technology to improve the algorithm and structure of Shack-Hartmann wavefront sensor and wavefront sensor based on interference,in order to improve the stability and detection accuracy.The main content of this dissertation can be divided into four parts:Firstly,the basic principle of adaptive optics system and some commonly-used wavefront sensors are introduced.The structure and algorithm principle of ShackHartmann wavefront sensor are analyzed in detail.It is pointed out that ShackHartmann wavefront sensor has the problem of insufficient robustness in extreme environment and the problem of insufficient measurement accuracy of high frequency information in high precision measurement environment.Through the analysis of phase extraction algorithm in interferometry,it is found that the existing phase-shifting interferometry algorithm needs too many interferograms and relies heavily on the accuracy of phase shifter.Then,the most commonly-used technologies of deep learning are summarized,and the application status of deep learning in adaptive optics wavefront detection and wavefront control is summarized in detail.From the existing research,it can be found that introducing deep learning technologies into wavefront sensing is highly feasible.However,there are also many problems at present,and has a huge room for improvement.In the second part,the Gaussian model of subaperture spot in low SNR and interference light environment is established,and the limitations of various improved Co G methods in extreme environment are analyzed.In order to solve the problem of closed-loop disorder caused by the failure of Shack-Hartmann wavefront sensor in extreme environment,a classification method of the stability of the deformable mirror in closed-loop based on logical regression is proposed,which can shut down the system in time when the closed-loop of adaptive optics system is abnormal to avoid the loss of system facilities.In order to make the adaptive optics system continue to work in extreme environment,the neural network computing graph of the improved Co G methods is analyzed in detail.It is found that all the improved Co G methods are special cases of the fully connected single hidden layer neural network.Therefore,a classification neural network SHNN is proposed,which can find out the pixels where the spot centroid is located,and a simulation training set containing 60000 data is generated to train the network.After the network is well-trained,the comparison between SHNN and traditional algorithms in simulations and real experiments show that the RMS of the residual phase restored by the optimal SHNN is nearly one order of magnitude smaller than that of the residual phase restored by the traditional threshold method.In the third part,four kinds of two-frame interferometry algorithms are introduced in detail,which can use two interferograms with unknown phase step to calculate the wrapped phase.Then the history of U-Net,a powerful tool in the field of computer vision,is summarized.The original U-Net is transformed to construct the Phase U-Net which can recover the wrapped phase from two interferograms.The data set is generated by careful simulations to train the Phase U-Net.Then,the performance of the Phase U-Net is simulated and analyzed in detail,and the principle and effectiveness of the neural network to calculate the wrapped phase are discussed.Compared with four traditional algorithms,the Phase U-Net algorithm is proved to be superior in accuracy.In the fourth part,in order to improve the ability of Shack-Hartmann wavefront sensor to detect sophisticated information of the phase,a defocused Shack-Hartmann wavefront sensor based on phase retrieval technology is proposed.Making full use of the spot shape information in subapertures to obtain the phase distribution information has been the direction of many researchers.However,how to obtain the phase information in subapertures from single intensity picture and how to fuse the phase information in subapertures with the slope information are always two difficulties to improve the accuracy of Shack-Hartmann wavefront sensor.In this dissertation,we first use the linear phase retrieval technique to recover the phase in each subaperture under the premise of small aberration.Then we construct the neural network LPR U-Net to fuse the results of the linear phase retrieval with the phase restored by the modal method.Simulation results show that the detection accuracy of this method is better than that of the classical Shack-Hartmann sensor.According to the specific requirements of wavefront sensors in different applications,this dissertation focuses on the algorithm research and experimental verification of adaptive optics wavefront sensing technology based on deep learning,which lays a foundation for further integration of deep learning technologies and adaptive optics system.