Deep Learning Based Prediction And Analysis For Light Fields Propagating Through Atmosphere And Optical Image Centroid Position

Due to light spot diviation and fields distortion caused by atmospheric turbulence in space optical communication,a photon flow model is used to describe the light field distributions and variations on CMOS to predict the position of the spot centroid;a fluid screen model is utilized to describe the light fields fluctuations,so as to analyze and predict the light fields distributions.In the view of theory,the complexity of light spot shape and fields distributions in atmospheric turbulence is due to the complexity of turbulent medium.Depending on the photon flow model and deep learning,the prediction of the spot centroid can make full use of the temporal correlation of light fields in turbulence.Based on the fluid screen model,the atmospheric channel,which is a typical open stochastic system,is now compressed into a semi closed two-dimensional fluid system.Thus,a traditional problem in statistical optics is transformed into a problem combined fluid fields prediction with classical optics.The complexity of the fluid is due to the interactions among micro clusters.In this paper,the relationship between fluid viscosity and spatial correlation of light field is investigated,and a model for predicting light fields based on fluid equation and Fourier optics is proposed.However,the traditional atmospheric optical methods still provide random boundary conditions for two-dimensional fluid screens,thus quantitative prediction is combined with statistical analysis.From the aspect of application,deep learning methods are fully utilized in this dissertation,a new neural network model is developed to deal with a large number of high-order tensors.Based on the fluid screen model and deep learning method,analysis and prediction methods without wavefront sensor are developed for light fields propagate in turbulence,which provides a complete theoretical model and an important experimental method for simplification on the adaptive optical system based on active correction.In this paper,deep learning network is a powerful fitting tool,and a reasonable physical model provides efficient constraints for deep learning optimization.According to this principle,a deep learning neural network model is developed with its strong nonlinear fitting ability,to reveal the essence of physical system.The purpose of this dissertation is to develop a self-consistent physical model,combined with deep learning method to provide a reliable model for describing laser propagation in atmosphere turbulence,so as to analyze and predict the light spot centroid position and light fields distributions.Experiments are carried out to verify the model,and eventually,to suppress effects of atmospheric disturbance.The two leading problems in this dissertation are as follows:(1)Based on deep learning and photon flow model,the position prediction model of optical image centroid is proposed.A relatively simple physical model coorporated with deep learning neural networks is developed to predict coordinates of the optical image spot centroid.The basic method is as follows: Image signals in the time series are taken as input,combined computer vision with deep learning,these image data are processed to generate physical information matrices,and then convolution neural network is applied to extract feature vectors containing position and motion information from physical information matrices.Then,the signals in time series composed of the feature vectors are input to the long-short term memory neural network to predict the position of the image spot centroid in future times.In this way,we can make full use of the temporal correlation of light fields in atmosphere turbulence to predict the position of spot centroid,so as to provide accordance for tracking and aiming process in space optical communication.Experiments are carried out to verify the prediction model.(2)Based on deep learning and fluid screen model,a relatively complex physical model combined with deep learning neural network is invented to analyze and predict light fields in atmosphere turbulence.The basic idea is,effects derived from atmospheric turbulence are assumed to be compressed into a sufficiently thin twodimensional fluid screen.In order to get the initial value of the fluid screen density distribution,a convolution neural network is used to process the optical images to fit current fluid screen density distributions,and a newly developed long-short term memory neural network is used to process these signals in time series to estimate velocity distributions in current fluid screen to obtain initial values for the fluid screen velocity distributions.Then,density and velocity distributions of fluid screen in future times are predicted by computational fluid dynamics.In this process,a small neural network is designed to deal with velocity gradients of the fluid screen to obtain viscous stress matrices.Information on density,velocity,refractive index distributions in atmosphere turbulence are contained in fluid screen.After the prediction on future state of fluid screen,light fields in future can be predicted with computational optics in a high efficiency.This provides a new method for prediction and calibration on the wavefront distortion for the active adaptive optical system.