Coherence- and scattering-universal diffractive neural network for imaging through scattering media

We propose a design method for a diffractive neural network (DNN) for imaging through scattering media, offering robustness against the spatial coherence of illumination, scattering strength, and scattering dynamics. Most techniques for imaging through scattering media are time-consuming and/or tailored to specific optical conditions. The DNN, composed of layers of diffractive optical elements (DOEs), optically reproduces the intensity distributions of objects behind scattering media without any computational processing. Datasets with randomized optical parameters are provided during the training process to achieve this robustness. We demonstrate the proposed method through numerical calculations and show its promising capability for DOE design. Our study paves the way for unifying and generalizing techniques for imaging through scattering media, which are currently fragmented by specific scenarios, enabling highly flexible imaging independent of optical conditions.