Computational Imaging in Digital Holographic Reconstruction with Machine Learning

We introduce a digital holographic reconstruction method using capsule-based deep learning network, which aims at overcoming information loss inherent in convolutional neural networks (CNNs). It takes into account the spatial relationship of neurons by embedding information in vectors instead of scalars. Experimental results demonstrate that capsule-based method is capable of reconstructing clear images from raw holograms without prior knowledge, and it produces comparable or even better performance than CNN-based holographic reconstruction method with much fewer parameters.