Physics-informed two-stage neural network enables computational ghost imaging through unknown scattering media

Imaging through scattering media located behind and in front of the object simultaneously remains a significant challenge in computational ghost imaging, as the reconstructed images are often severely degraded due to distorted illumination patterns and attenuated intensity values. In this Letter, we propose an untrained, two-stage self-supervised neural network that effectively mitigates both forward and backward scattering effects. Without the need for any labeled training data, our physics-informed framework exhibits strong adaptability to various types of unknown scattering media. We experimentally demonstrate the robustness of the proposed method by imaging through different scattering environments, including biologically relevant media and rotating ground glass. Compared with the existing reconstruction algorithms, our approach achieves substantially improved image quality for computational ghost imaging in unknown scattering media. This study introduces a paradigm for imaging through complex media and paves the way toward practical applications in biomedical and remote sensing scenarios.