Research On Image-based Wavefront Sensing Technique Using Deep LSTM Network With Phase Diversity Images

Image-based wavefront sensing methods have the advantages of simple hardware implementation and high wavefront measurement accuracy,and therefore they have good application prospects in the fields of adaptive optics and active optics.However,the current image-based wavefront sensing methods(including phase retrieval and phase diversity algorithms)need to use iterative or numerical optimization to solve the wanvefront aberrations,which has many iterations and takes a long time to converge.Moreover,with the increase of the range and dimension of the aberration coefficient,the probaBility of falling into the local optimum greatly increases,which seriously affects the wavefront sensing accuracy and restricts the application scope of imagebased wavefront sensing.At present,deep learning has been widely used in pattern recognition,data mining and other fields.Compared with the traditional iterative algorithms such as phase recovery and phase diversity,it has the potential advantages of fast computing speed(no iterative process),high accuracy(no local minimun problem)and so on.However,deep learning is rarely used in the field of image-based wavefront sensing.It is necessary to explore and study the deep learning model in order to improve the accuracy and practicaBility of the deep learning method in the field of wavefront sensing.This paper mainly includes the following parts:Aiming at the problem of wavefront sensing using point target image information,this paper proposes a deep learning model based on convolution neural network and deep LSTM network respectively,which can use two PSF images with fixed phase diversity to solve the aberration coefficients.Simulation and experimental results show that the deep LSTM neural network has high accuracy and calculation efficiency.Aiming at the problem of wavefront sensing using arBitrary target image information,this paper first proposes a method of feature image extraction,which extracts feature of two images with phase diversity in frequency domain.It can obtain a feature image independent of target information and only related to the wavefront phase information of optical system.On this basis,a deep learning model based on deep LSTM network is proposed to solve the wavefront phase from a pair of arBitrary target images with phase diversity.Aiming at the noise problem,this paper proposes a denoising method based on N2 N denoising theory and deep convolution neural network.In practice,the image can be denoised first,and then input it into the depth LSTM network to solve the wavefront phase information.The effectiveness of the proposed method is verified by simulation and experiment.Aiming at the problem of wavefront sensing using arBitrary target image information with unknown bandwidth spectrum and spectral distribution,this paper first proposes a method of feature image extraction based on matrix theory,which uses two images with phase diversity to obtain a feature independent of the spectral distribution of light source and target information,and related to the wavefront phase information and spectral range of optical system image.On this basis,a deep learning model based on the deep LSTM network is proposed to calculate the wavefront phase information and recover the target image using the extracted feature image information.The simulation and experiment results show that the proposed model can accurately calculate the wavefront phase information and restore the clear original target image.This paper solves many problems in the field of image-based wavefront sensing,which is of great significance to improve the performance of image-based wavefront sensing.