Phase retrieval of color images based on a self-supervised deep neural network

Abstract

Phase retrieval is one of the vital processes in many computational imaging techniques, aiming at retrieving the lost phase of light fields from the corresponding intensity. Recently, advanced deep learning strategies for phase retrieval have gained much attention mainly due to their highly efficiency and accuracy compared with conventional iterative methods represented by Gerchberg-Saxton algorithm. Here, we propose a self-supervised neural network integrated with angular spectrum transform to retrieve lost phases of color images. The network contains two complex-valued U-Nets to restore and update the light fields of the image plane and the object plane, respectively. By minimizing the difference between the reconstructed images from the network with the input images, complex light fields on the object plane can be obtained. The proposed network is able to retrieve missing phases of 200 color images within half a minute while the averaged peak signal-to-noise ratio and structural similarity of the reconstructed color images can reach 23.21 dB and 0.84, respectively. Visualization and statistic results indicate that our network is an efficient and accurate method of phase retrieval, which has potential applications in many fields of computational imaging.