Unpaired learning for digital holographic reconstruction and generation

Traditional numerical reconstruction methods in digital holography (DH) are faced with problems such as inaccurate and time-consuming unwrapping or the need to capture multiple holograms with different diffraction distances. In recent years, deep learning, believed to be a new and effective optimization tool, has been widely used in digital holography. However, most supervised deep learning methods require large-scale paired data, and their preparation is time-consuming and laborious. Here, we propose what we believe to be a new deep learning approach that can use less unpaired data to train neural networks, thereby reducing the need for labeled data. By using CycleGAN, the calculation process of the loss function does not need paired data, and the network can learn the mapping knowledge between two domains, instead of the data and its corresponding label. As such, the proposed method can reconstruct complex amplitudes for holographic reconstruction. And more importantly, the trained network learns the knowledge of the imaging system, leading to the capability to generate experimentally synthetic holograms at the same time. To the best of our knowledge, this is the first demonstration of such high-quality physics-based experimental hologram simulator in DH. The reconstructed complex amplitudes have higher image quality, while the generated holograms can represent realistic system status and reconstruct the input complex amplitudes.