Physics-constrained Mamba-UNet for enhanced-resolution single multimode fiber ghost imaging at low sampling rates

Single multimode fiber (MMF) ghost imaging holds significant potential for endoscopic applications; however, its practical implementation is hindered by severe mode coupling and the ill-posed nature of reconstruction under low sampling rates. To address these challenges, we propose a high-fidelity reconstruction framework based on a Physics-Constrained Mamba-UNet. Unlike conventional convolutional networks restricted by local receptive fields, our architecture leverages a visual state space model (VSSM) to efficiently capture long-range spatial dependencies within speckle fields while maintaining linear computational complexity. Furthermore, we explicitly embed a differentiable forward physical observation model into the training process, transforming the optimization from pure data fitting into a physics-constrained inverse problem. Both numerical simulations and experimental results in a single-MMF system demonstrate superior reconstruction performance, achieving a 2.2-fold resolution enhancement and high structural similarity even at sampling rates as low as 5%. This framework significantly improves robustness against mode-mixing noise, offering a viable pathway for real-time, high-resolution fiber endoscopy.