Research On Phase Retrieval Algorithm Based On Prior Information

The original signal is reconstructed in phase retrieval only from intensity or amplitude measurements,and phase retrieval are widely used in computational imaging systems,such as lens free imaging,scattering medium imaging and coherent diffraction imaging.In the light field,the signal consists of two parts: amplitude and phase,while the phase contains the main information of the signal.Owing to the insufficient sampling frequency,the existing photosensitive sensors such as charge coupled devices can only record the signal intensity information of the signal,without the phase information of the signal.Using only intensity information to solve the phase retrieval model can obtain multiple solutions,so phase retrieval is an ill-posed problem.Therefore,prior information needs to be used to ensure the accurate reconstruction of the original signal.Because of the low quality of the reconstruction signal with the measured value containing a lot of noise,the denoising regularization framework is added into the phase retrieval model in this thesis,and noise is used as a prior information to constrain the original signal reconstruction.And the prior information will improve the robustness of the phase retrieval algorithm.For the problem of low resolution of image reconstructed in phase retrieval,the super resolution prior is applied to the phase retrieval model.And the adaptive gradient descent method is selected to solve the model to retrieval the high resolution original image.This thesis mainly considers to use prior information to improve the reconstruction quality of phase retrieval algorithm.The specific research contents are as follows:(1)In practice,the measurement value is inevitably disturbed by noise.Thus this thesis adds the convolutional neural network Dn CNN into the denoising regularization framework as a noise operator,and provides the phase retrieval algorithm Nr PR＿Dn CNN which based on the convolutional neural network.The provided algorithm decomposes the phase retrieval into two subproblems: image denoising and constraint optimization.Dn CNN is used to remove noise in the image denoising subproblem,then the constraint optimization subproblem is solved by rapid gradient descent method.The two subproblems are solved alternately until the stop condition is reached.Simulation experiments show that the proposed Nr PR＿Dn CNN algorithm can reconstruct the original image with small error at different noise levels,and this algorithm has convergence.(2)Classical phase retrieval algorithm is limited by the resolution of the imaging system,result in the low quality reconstructed image.To address this problem,this thesis combines the super resolution reconstruction with phase retrieval algorithms,and provides the super resolution phase retrieval algorithm based on single and multiple frames.In the single frame super resolution phase retrieval algorithm,the accurate initialization value is obtained using the truncated initialization method.And in the step selection experiments,the adaptive gradient descent method with the smallest iteration error was selected to solve the single frame super resolution phase retrieval model.By exploring the effect of different magnification in the retrieval performance of the algorithm,it’s known that the greater the magnification,the higher the image reconstruction quality.Owing to the single frame algorithm contains less image potential information in the amplitude measurement value,the denoising algorithm is easy to loss the edge and texture details of the image.To solve this problem,this thesis provides a multiple frame super resolution phase retrieval algorithm to improve the reconstruction performance by obtaining more potential information.Single frames is a special case of multiple frames,and the multi-frame algorithm reconstructed high-resolution original images by sparse initialization and gradient descent methods.Simulation experiments with different frames show that the more frames are sampled,the better the reconstruction effect.At the same time,the reported algorithm can effectively suppress noise and visual artifacts.Moreover,compared with the classical super resolution phase retrieval algorithm at multiple noise levels,the provided algorithm has the best denoising effect and the highest reconstruction quality.which verifies the robustness and effectiveness of the proposed algorithm.