Phase Retrieval Based on Enhanced Generator Conditional Generative Adversarial Network

Abstract

Phase retrieval refers to the recovery of the original image using only the Fourier amplitude of the image. Due to the small amount of information contained in the Fourier amplitude, the common network structure cannot achieve accurate reconstruction of the image when the oversampling rate of the image is low. It is the key issue of phase retrieval to improve the structure of the neural network. We propose an application of end-to-end adversarial network to solve phase retrieval problems by adding a U-Net model to the conditional generative adversarial network(U-NetCGAN). This desired approach realizes multi-scale recognition and fusion of image features and improves the quality of image reconstruction. The experimental results show that the model is significantly better than the traditional phase retrieval algorithm. Compared to other algorithms, the evaluation indicators of PSNR and SSIM values in our approach have increased about 6 dB and 0.1, respectively.