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

Area of Science:

• Optics and Photonics
• Computer Vision
• Astrophysics

Background:

• Atmospheric turbulence significantly degrades image quality in telescopic observations.
• Single-pixel imaging (SPI) offers robust multi-wavelength capabilities but requires advanced reconstruction techniques.
• Deep learning (DL) presents a promising approach for overcoming imaging challenges in dynamic random media.

Purpose of the Study:

• To integrate deep learning into a single-pixel imaging system for improved telescopic observation under atmospheric turbulence.
• To develop and evaluate a single-pixel telescope system capable of simultaneous visible and near-infrared (NIR) imaging.
• To compare the performance of U-Net and time-division pattern learning (TDPL) networks for image reconstruction in simulated turbulent conditions.

Main Methods:

• Development of a single-pixel telescope system for simultaneous visible and NIR observation.
• Numerical simulation of atmospheric turbulence with dynamic phase disturbances.
• Comparative performance evaluation of U-Net (non-temporal) and TDPL (temporal) deep learning networks for image reconstruction.

Main Results:

• The U-Net model demonstrated higher accuracy in reconstructing simple targets (e.g., MNIST images) compared to the TDPL network under simulated turbulence.
• The TDPL network, while designed to capture temporal fluctuations, did not outperform U-Net in this specific experimental setup.
• The study confirmed the potential of fusing SPI's multi-wavelength capability with DL-based noise suppression for robust imaging.

Conclusions:

• Highly optimized deep learning architectures like U-Net are currently a practical and effective strategy for single-pixel imaging reconstruction tasks.
• Further refinement of temporal learning methods like TDPL may enhance future performance.
• The integration of SPI and DL offers a resilient and extensible imaging technology for challenging observational scenarios.