Research On Wavefront Correction Method Of Convolutional Neural Network Based On Bayesian Optimization

The wavefront of a propagating beam is prone to be disrupted by the random medium,the Adaptive Optics(AO)technology is used to correct the wavefront aberrations,and eliminate its influence on the imaging system.However,when the wavefront aberration changes rapidly,traditional algorithms usually suffer from long calculation time and limitation of environment,and consiqaently cannot meet the requirements of the adaptive optics system for real-time correction of the wavefront.Therefore,this paper proposes a non-iterative wavefront correction method based on Convolutional Neural Network(CNN)under Bayesian Optimization(BO)to ompensate the wavefront in real time.Firstly,in order to combine CNN with adaptive optics,the method of generating intensity pattern of distortion employs Kolmogorov spectrum to generate random coefficients of Zernike polynomials to fit wavefront aberrations.Based on this,CNN uses the intensity pattern as training samples,and then trians in model for predicting Zernike coefficients by Adaptive moment estimation(Adam)and Dropout methods which deploy the mean square error loss function.Second,to seek a suitable prediction model of CNN,the model selection and performance verification are exerted under Bayesian optimization through the Python platform and Tensorflow framework for numerical simulation.The BO method is used to adjust the model twice to obtain the three-layer network parameters and training parameters that will converge within200 iterations.According to the parameters,16000 sets of samples are used to train the network,and the Convolutional Neural Network of three layers under Bayesian Optimization(CNNUBO3)model with the loss function reduced to 0.1 is obtained.By comparing CNNUBO3 with Stochastic Parallel Gradient Descent Algorithm(SPGD)and genetic algorithm,it is found that CNNUBO3 is significantly better than SPGD and genetic algorithm in weak turbulence,the average correction time is decreased to 1ms,and the performance indicators are improved.Finally,in order to verify the application effect of CNNUBO3 in the wavefrontsensorless adaptive optics system,an experimental scheme of non-iterative wavefront correction is designed.When simulating turbulence,The phase screen changes the optical path according to different refractive indexes and consiquently disturbs the wavefront.the deformable mirror generates the conjugated wavefront,which can offset the optical path difference introduced by the phase screen.Therefore,to control the piezoelectric deformable mirror,a series of Zernike coefficients obtained by CNNUBO3 according to the intensity pattern are delivered into the control program of deformable mirror.The experimental results show that after applying CNNUBO3 to wavefront correction,the strehl ratio and the image sharpness function change significantly.Correction time can be decreased to 1.6ms.Simulations and experiments prove that the method of wavefront correction based on CNN and BO is feasible,and has a certain improvement in response time and performance compared with traditional methods.