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

Ghost imaging (GI) reconstructs objects using single-pixel measurements and is widely explored for remote sensing, imaging through scattering media, and photon-limited environments. However, deep-learning-based computational ghost imaging (CGI) often relies on experimentally acquired low-resolution datasets, making data collection time-consuming and limiting denoising performance. This work reports on a simulation-based training strategy that generates high-resolution synthetic datasets replicating experimental conditions, enabling efficient network training without extensive data acquisition. Using this approach, the convolutional blind denoising network (CBDNet) achieved peak signal-to-noise ratio (PSNR) values up to 12.79 dB for complex experimental targets and approximately 10.5 dB for structured targets at 256 × 256 resolution, while preserving fine details in cross-sectional intensity profiles. These results demonstrate that simulation-driven training significantly enhances denoising performance and scalability, paving the way for high-resolution ghost imaging in complex and photon-starved scenarios.