Learning‐Based Vectorial Reconstruction of Orthogonal Polarization Components in a Structured Vector Optical Field Passing Through Scattering Media

Abstract Optical imaging through scattering media has become important due to its fundamental physics interest and various applications. The reconstruction of a structured optical field with various states of polarization passing through a scattering medium with a speckle pattern behind the scattering medium remains challenging since existing restoring techniques only reconstruct the speckle in a single‐polarization state (scalar optical field). This work proposes a novel approach to simultaneously restore the initial orthogonally polarized components from a speckle pattern behind a scattering medium. The neural network Polarization‐DenseUnet (P‐DenseUnet) based on the vector transfer matrix is constructed to restore the two orthogonally linear (or circular) polarization components of a structured vector optical field from a speckle pattern behind the scattering medium. The generalization and effectiveness of this proposed method are tested for high fidelity with different phase distributions such as vortex, digits, and Fashion‐mnist.