Imaging with a single-pixel detector is challenging, when an object is hidden in time-varying scattering media. The correlation between illumination patterns and collected single-pixel light intensities is severely destroyed. In this Letter, we report high-fidelity correspondence imaging of a hidden object in time-varying scattering media. A real-valued noiselet matrix with orthonormal columns is generated and applied to offer high robustness against scattering media. To eliminate dynamic scaling factors, the collected single-pixel light intensities are corrected by using a sequence of estimated expectations. Then, an untrained neural network is designed and fine-tuned to recover a high-quality object image without any datasets and labels. It is validated in experiments that our method can always achieve high-quality reconstruction of an object hidden in time-varying scattering media with light disturbances. This work could pave the way for high-quality imaging of an object hidden in complex media.
Xu et al. (Mon,) studied this question.