Large-scale phase retrieval

Abstract High-throughput computational imaging requires efficient processing algorithms to retrieve multi-dimensional and multi-scale information. In phase imaging, retrieval (PR) is required reconstruct both amplitude in complex space from intensity-only measurements. The existing PR suffer the tradeoff among low complexity, robustness measurement noise strong generalization on different modalities. this work, we report an large-scale technique termed as LPR . It extends plug-and-play generalized-alternating-projection framework real nonlinear space. alternating projection solver enhancing neural network are respectively derived tackle formation statistical prior regularization. This compensates shortcomings of each operator, so realize high-fidelity with complexity generalization. We applied for a series modalities including coherent diffraction coded pattern Fourier ptychographic microscopy. Extensive simulations experiments validate that outperforms much 17dB enhancement signal-to-noise ratio, more than one order-of-magnitude increased running efficiency. Besides, first time demonstrate ultra-large-scale at 8K level ( $$7680\times 4320$$ 7680 × 4320 pixels) minute-level time.