Learning based holographic reconstruction through a diffuser

Abstract

Object recovery from speckle patterns has been extensively studied, and holographic reconstruction technique is verified to be highly effective for recovering objects. However, for the holograms recorded through scattering media, conventional holographic techniques cannot retrieve useful information from the holograms. In this paper, we present an approach based on convolution neural network (CNN) for holographic reconstruction through a diffuser. The object is placed behind a diffuser, and the corresponding hologram is recorded by a CCD camera. With pairs of holograms and their original objects sent to the designed learning structure, a CNN model is trained to perform unknown-object retrieval from the holograms. This learning-based approach can make predictions of the unknown test objects in real time. It provides a feasible structure to conduct object recovery from the holograms recorded through a diffuser.