Fourier phase retrieval using multiple constraints based on physics enhanced neural network

Recovering an object only from the amplitude of its Fourier measurement is a long-standing challenge. To confront this intricate challenge of illness more effectively, we propose a framework that combines data-driven pre-training and physics-driven iteration. These constraints including adapted support region and noise of image, which comes from the feature of object itself. Our analysis of both simulated and optical experiments data reveals that this framework offers superior results than other methods. Moreover, this improvement is achieved without suffering from the limitation of the dataset, may cast new light on network based algorithm in the future.