Phase retrieval from a single diffraction intensity pattern by generating the support constraint using deep learning

Abstract

Phase retrieval is an essential technique for imaging applications in many fields such as biology and materials science. The fundamental challenge lies in reconstructing objects from measured intensity patterns. A prevalent solution involves an iterative strategy that alternates projections between the object and detector planes. However, reconstructing objects from a single diffraction intensity pattern remains challenging due to the absence of constraints at the object plane. Here, we propose a learning-based approach that generates reliable object support in real time from its autocorrelation support. With the generated support as the constraint for the object plane, we show that the convergence of iterative phase retrieval algorithms can be markedly enhanced. Compared with the end-to-end learning-based methods, the proposed method has a unique character that the training features are binary supports, instead of the type and content of objects, so that our approach has better generalization. Furthermore, one of the advantages of our method is that it uses simulated data for training, eliminating the need for experimental data collection.