The Research Of Phase Retrieval Algorithm Based On Deep Priors

The problem of phase retrieval is recovery of a signal only from the magnitude of its Fourier transform,or of any other linear transform.Due to the loss of phase information,this problem is ill-posed.Therefore,the prior knowledge is required to enable its accurate reconstruction.For the various phase retrieval algorithm based on regularizations has the disadvantages of low reconstruction quality and high computational complexity under the presence of noise.We combine traditional phase recovery algorithms with convolutional neural networks in order to improve the above problems.The main contents are as follows:Firstly,for the practical application,the reconstructed image is often subject to noise interference.We use the residual learning of deep convolutional neural network?DnCNN?model which is trained for removing image noise as a denoiser.At the same time,the l₂norm model is introduced to limit the solution space,and controls the complexity of the model.We propose to a phase retrieval algorithm based on DnCNN-l₂ norm.This model is non-convex,and the alternating direction method of multipliers?ADMM?is utilized for solving the corres-ponding non-convex optimization problem.At last,verify the effectiveness of the proposed algorithm through experiments.Secondly,in order to improve the l₂ norm model is added as regularization into the phase retrieval problem has high computational complexity and solving the defect of slow recovery speed,we proposed hybrid input-output?HIO?phase retrieval algorithm based on encoder-decoder network.In the iterative process of HIO algorithm,the reconstruction quality of HIO's intermediate results is improved by encoding-decoding network.The experimental results show that the proposed algorithm has noise robustness.Finally,since the HIO algorithm still has a lot of limitations on simulated data with varying amounts of Poisson noise,we proposed the oversampling smoothness?OSS?phase retrieval algorithm based on deep image prior.We use the contraction-expanded convolutional neural network as a deep prior.The algorithm taking the deep prior?support constraints and amplitude constraints together as the prior of the OSS algorithm.After many experiments,it proved that the proposed algorithm can not only restore more image details,but also show more robustness to noise.