Design And Implementation Of Fast And Sensitive Wavefront Sensor Based On Deep Learning

As an efficient wavefront detector,Shack-Hartmann is often used as a wavefront phase detection device in adaptive optics(AO)systems.The combination of AO and microscopy technology can restore diffraction limited imaging of biological tissue by using active elements such as DM or SLM,therefore obtaining high-resolution image of biological tissue.This requires to improve the speed and sensitivity of traditional Shack-Hartmann wavefront detector.In recent years,researchers have greatly improved the detection speed and sensitivity of traditional Shack-Hartmann wavefront detector through the continuous efforts in algorithm and hardware,However difficulties still exist in the application of two-photon microscope.Based on the summary of centroid calculation and prediction algorithm of Shack-Hartmann wavefront detector,the deep convolution neural network is introduced into Shack-Hartmann wavefront detector,and a deep learning framework of convolution neural network with VGG-Net as the backbone is proposed.The centroid position information is extracted from dot matrix by convolution layer.The activation function between the convolutional layers introduces non-linearity during the centroid information extraction process and plays a role in suppressing noise.The fully connected layer after convolution regresses the wavefront coefficients.The centroid calculation and the wavefront coefficients prediction process are integrated into the same framework,and an end-to-end wavefront detection algorithm is completed to achieve a fast and sensitive Shack-Hartmann wavefront detector.Through the reasonable design of the network,the network is first trained on the simulated data,then fine-tuning the network with the real data.Finally,a good prediction result was obtained on the test data set containing simulated tissue noise,and the root mean square error of the true value and the predicted value reached 0.06?.In this paper,the feasibility of the method is also verified by an ideal AO system.In the real AO system,a small spherical aberration is corrected,and the feasibility of this method in practical situations is analyzed by comparing the point spread function before and after correction.