Differential ghost imaging with learned modulation patterns

Unlike conventional imaging with two-dimensional array sensors featuring millions of pixels, ghost imaging enables the use of advanced detector technologies, giving advantages such as high signal-to-noise ratio, wide spectral range, and robustness to light scattering. However, this involves an extremely time-consuming measurement process, which means that it is difficult to meet the needs of high-quality real-time imaging. This paradox becomes notable especially in the context of utilizing non-orthogonal modulation patterns, such as the speckles generated by rotating ground glass. Efficient modulation patterns and advanced reconstruction algorithms are widely studied as two main ideas to solve the above problem. Here, we perform real-time, high-fidelity differential ghost imaging (DGI) at a low sampling ratio of 6.25% by proposing a compact physically guided single-layer neural network with the DGI algorithm embedded. Simulations and experiments show that, once the learned modulation patterns are obtained, our scheme can achieve fast, high-quality, and noise-robust DGI without the need for complex iterative optimization algorithms or subsequent optimization neural networks. Our scheme opens up new horizons for exploring more efficient modulation patterns for ghost imaging by deeply combining physical priors.