A Deep‐Learning Approach for Low‐Spatial‐Coherence Imaging in Computer‐Generated Holography
Xin Tong, Renjun Xu, Ke Liu, Liangliang Zhao, Weilai Zhu, Daomu Zhao
Abstract
The low‐spatial‐coherence imaging capability of computer‐generated holography (CGH) is a key to high‐resolution displays, virtual reality, augmented reality, and holographic microscopy. The low spatial coherence caused by complex disturbances can damage the image quality irreversibly. The optical field with low spatial coherence has large fluctuations, making it difficult to be quantified and modeled directly. To tackle these challenges, a deep neural network‐based model U‐residual dense network (U‐RDN) is proposed, which obtains the optimal solution under the low‐spatial‐coherence condition. The large‐scale images are generated using optical experiments, with analysis and restoration by deep learning. Extensive experiments demonstrate the strong out‐of‐distribution robustness of U‐RDN, which is generalizable to unseen classes in unseen domains. The learning‐based approach and low‐spatial‐coherence dataset open a new path toward the next generation of CGH.