Gaussian priors and physics-informed for domain-agnostic ghost imaging

Ghost imaging transcends the constraints of classical optical imaging by coupling low-dimensional encoding with high-dimensional decoding. Existing methods rely heavily on prior modeling constraints and specific data, making it difficult to balance reconstruction accuracy, computational speed, and system adaptability. Here, we demonstrate a unified framework for ghost imaging that combines Gaussian priors with the physics-informed method. Gaussian image pre-training enables domain-agnostic regularization to be embedded into models, which can mitigate reliance on large-scale natural image datasets and decouple data generation from reconstruction characteristics, enhancing its generalization capability. The physics-informed fine-tuning stage is attuned to the specific system sampling conditions, culminating in high-fidelity image reconstruction. Simultaneously, we design a lightweight neural network, LGINet, tailored for ghost imaging. While preserving model performance, it reduces computational overhead, lowers storage demands, and accelerates convergence. The presented paradigm of synergistic ghost imaging features strong generalization, expeditious reconstruction, and high-quality imaging, which will bring new access to efficient imaging in the background of resource-constrained or heterogeneous requirements.