The Research Of Phase Retrieval Algorithm Under Incomplete Diffraction Pattern Information

Phase retrieval(PR)refers to recover the original signal by using the intensity measurement of its Fourier transform,or other linear transform.In the lens less coherent diffractive imaging(CDI)experiments,a beam stop is often used to block strong intensities which exceeds the limited dynamic range of the sensor.Therefore,the information in the center region of the diffraction pattern,namely the low-frequencies information,is lost.Recovering the image with high quality only from the incomplete diffraction pattern information is a challenge.To address this issue,in this work,we explore the sparse information of image signals under different gradient sparsity operators as prior and explore how to improve the image quality from the incomplete diffraction pattern.The research contents are as follows.First,combined with the idea of nonlinear compressed sensing,using the first derivative of the gradient domain image as the regularization term and combining the support constraint in real space,the paper proposes an iterative phase retrieval framework.Experimental results indicate that the proposed algorithm can reconstruct images better and robust to the different missing size in the diffraction pattern.Secondly,considering the real-world images have shown the marginal distributions well modeled by a hyper-Laplacian.The paper proposed method of fractional variation as the regularization term.Moreover,the alternating direction method of multipliers(ADMM)is utilized for solving the corresponding non-convex optimization problem.Experimental results show that the performance of this algorithm outperforms the HIO+CS at the most cases.Finally,the TV only calculates the first derivative of the image,therefore the smooth of the high-order is not take into account.The paper proposed a phase algorithm based on the total generalized variation(TGV)regularization.Due to the Poisson noise existed in the CDI,a Poisson data fidelity term that derives from the maximum-likelihood method is utilized.The data fidelity term,the TGV regularization term and the support constraint term are fused to formulate a PR optimization problem.The experiment shows that this proposed algorithm outperforms the HIO+CS algorithm and the phase retrieval toolbox Hawk.