Hyperspectral dual-comb compressive ghost imaging with deep learning reconstruction

Abstract Ghost imaging reframes image formation by replacing pixelated sensors with correlation measurements between structured illumination and a single-pixel detector, enabling capabilities that are difficult or impossible for conventional cameras. This powerful technique is inherently suited to imaging through scattering media, photon-starved scenes, and size-, access-, or cost-constrained platforms. Yet its broader adoption has been limited by a fundamental speed bottleneck: slow, sequential pattern projection via spatial light modulators or wavelength sweeps, which precludes real-time capture of dynamic scenes. Here, we overcome this fundamental bottleneck by combining dual-comb interferometry with compressive ghost imaging and deep learning reconstruction. Our method, hyperspectral dual-comb compressive ghost imaging, utilizes optical frequency combs to create wavelength-multiplexed speckle patterns that are delivered through a single-core fiber and detected by a single-pixel detector. This approach enables the simultaneous illumination and detection of hyperspectral speckle patterns without slow spatial light modulation or wavelength scanning, thereby allowing snapshot compression and the acquisition of image information using spatially non-resolved hardware. To decode the compressed signals, we develop a transformer-based deep-learning model capable of rapid, high-fidelity image reconstruction at sampling ratios below 1%. Our hardware–software co-design approach achieves high-fidelity dynamic ghost imaging at video rates and beyond while dramatically simplifying the optical front end, significantly advancing ghost imaging toward practical deployment in size- and access-constrained scenarios—from minimally invasive endomicroscopy to compact portable sensors.