Phase retrieval from two diffraction patterns with coherent modulation using physics-driven neural network

Phase retrieval is a valuable technique in many imaging fields. While data-driven deep learning methods have demonstrated their potential in phase retrieval, they are often limited by the availability of data and their generalization capabilities. Physics-driven deep learning approaches offer an alternative strategy for phase retrieval, integrating a physical model with a neural network to eliminate the need for substantial amounts of labeled data. Here, we propose a dual modulation physics-driven deep learning scheme to improve the performance of phase retrieval. In the proposed method, two modulated diffraction patterns are acquired by using the different wavefront modulators. Then, an untrained neural network is employed to predict the objects by combining with the modulated diffraction process. By applying dual modulation constraints, we demonstrate that the neural network can learn phase retrieval more efficiently. Numerical simulations are conducted to verify the effectiveness and superiority of the proposed method.