High-resolution ghost imaging denoising using simulation-based deep learning.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Ghost imaging (GI) reconstructs objects using single-pixel measurements, valuable for remote sensing and imaging through scattering media.
• Deep-learning-based computational ghost imaging (CGI) typically uses low-resolution experimental data, leading to long acquisition times and suboptimal denoising.

Purpose of the Study:

• To develop an efficient training strategy for computational ghost imaging (CGI) using simulated high-resolution datasets.
• To improve the denoising performance and scalability of CGI networks without extensive experimental data collection.

Main Methods:

• A simulation-based training strategy was employed to generate high-resolution synthetic datasets that mimic experimental conditions.
• The convolutional blind denoising network (CBDNet) was trained on these synthetic datasets for enhanced denoising capabilities.

Main Results:

• The simulation-driven CBDNet achieved significant peak signal-to-noise ratio (PSNR) improvements: up to 12.79 dB for complex targets and ~10.5 dB for structured targets at 256x256 resolution.
• The network successfully preserved fine details in cross-sectional intensity profiles, demonstrating effective denoising.

Conclusions:

• Simulation-driven training substantially enhances denoising performance and scalability in ghost imaging.
• This approach enables high-resolution ghost imaging in complex and photon-limited scenarios, overcoming limitations of experimental data acquisition.