Unsupervised Deep Learning Enables 3D Imaging for Single-Shot Incoherent Holography

Incoherent digital holography no longer requires spatial coherence of the light field, breaking through the imaging resolution of coherent digital holography. However, traditional reconstruction methods cannot avoid the inherent contradiction between temporal resolution and signal-to-noise ratio, which is mitigated by deep learning methods, and there are problems such as dataset labeling and insufficient generalization ability. Here, a self-calibrating reconstruction approach with an untrained network is proposed by fusing the plug-and-play nonlinear reconstruction block, the forward physics imaging model, and a physically enhanced neural network. Measurement consistency and total variation kernel function regularization are used to optimize the network parameters and invert the potential process. The results show that the proposed method can achieve high fidelity, high signal-to-noise ratio, dynamic, and artifact-free 3D reconstruction using a single hologram without the need for datasets or labels. In addition, the peak signal-to-noise ratio of the reconstructed image with the proposed method is improved by a factor of 4.6 compared to the previous methods. The proposed method leads to considerable performance improvement on the imaging inverse problem, bringing new enlightenment for high-precision unsupervised incoherent digital holographic 3D imaging.