Mueller matrix-based deep neural network for adaptive reconstruction of orthogonal polarization components of a vector beam transmitted through a highly scattering medium

Imaging reconstruction through a highly scattering medium (HSM) has emerged as a hot research topic due to fundamental physics and significant potential applications such as biomedical imaging, optical testing and light-material interaction. Current deep learning reconstruction methods from a speckle behind a HSM have demonstrated their efficiency and flexibility. However, their dependence on large datasets and the lack of physical interpretability remain significant limitations. Here a novel physical-driven deep neural network is proposed that combines Mueller transformation matrix (MTM) with deep learning for scattering reconstruction from a speckle pattern. The MTM-based reconstruction network (MTMRNet) is designed by leveraging MTM to describe the polarization information evolution of scattered light fields passing through a HSM, combined with the strong feature learning and mapping capability of deep neural network. The experimental reconstruction results demonstrate the superior performance of the proposed lightweight neural network, offering a concise and reliable approach for reconstructing imaging from a speckle behind an (isotropic or anisotropic) HSM with reduced computing cost. The computing consumption of MTMRNet is nearly one order of magnitude less than that of Trans-CNN. These results can promote the transformation from an “data-driven” model to a new “physics-data dual-driven” model for deep learning-based reconstruction with a speckle behind a HSM.