High-fidelity ghost imaging using untrained physics-driven dual-network framework at ultra-low sampling rates

Abstract

Ghost imaging (GI) is a typical computational imaging technique that reconstructs two-dimensional and three-dimensional images from one-dimensional bucket detector signals under structured light illumination. By utilizing single-pixel detection, this technology is particularly advantageous in low-light environments and in spectral regions (e.g., infrared, ultraviolet, or x-ray), where high-performance array detectors are often impractical or prohibitively expensive. However, traditional GI methods suffer from poor image reconstruction quality at low sampling rates and high hardware requirements; additionally, the generalization issues of data-driven deep learning methods limit their practical applications. Here, we propose a physics-driven Dual Untrained Ghost Imaging Neural Network (DUGIN). By integrating a “coarse-to-fine” dual-network architecture with the physical model, our method utilizes the deep image prior to achieve stable optimization and effectively escape local optima. Furthermore, a general affine scale correction module is designed to compensate for the intensity scale bias caused by normalization, further improving reconstruction fidelity. Simulation and experimental results demonstrate that DUGIN achieves high-fidelity natural image reconstruction at a 5% sampling rate, showing significantly reduced image noise and clearer details compared to traditional differential ghost imaging and recent physics-driven Ghost Imaging using Deep Neural Network Constraint methods. This study provides a novel framework for GI technology and paves the way for its practical application.