Deep-learning-based single-pixel telescope for simultaneous visible and near-infrared imaging with robustness to atmospheric seeing

Attaining high-quality imaging through dynamic random media, such as atmospheric turbulence, poses a significant challenge in various fields of telescopic observation. This study addresses the challenge by integrating a deep learning (DL) technique into a single-pixel imaging (SPI) system known for its robust multi-wavelength imaging capabilities. We developed a single-pixel telescope system for simultaneous visible and near-infrared (NIR) observation and evaluated its performance under numerically simulated atmospheric turbulence. The simulation environment mimicked realistic atmospheric conditions with changing phase disturbances over space and time. We compared the performance of the widely used image-to-image model U-Net, which does not utilize temporal information, with our previously proposed time-division pattern learning (TDPL) network, which retains temporal structure in a signal-to-image framework. While TDPL holds potential advantages in extracting fluctuation characteristics from time-varying measurements, experimental results under numerically simulated turbulence conditions showed that U-Net achieved higher accuracy for simple targets such as MNIST images. Although further refinement of TDPL may change the situation in the future, leveraging highly optimized architecture such as U-Net represents a practical and effective strategy in SPI reconstruction tasks at present. The fusion of SPI's multi-wavelength capability with DL-based noise suppression provides robust and precise imaging under realistic observation conditions, highlighting its potential as a resilient and extensible imaging technology.