Higher-Order Modes Are More Robust: The Origin of Bending Insensitivity in Multimode Fibers and Its Exploitation for Imaging
Multimode fibers (MMFs) hold great promise for minimally invasive imaging, yet their extreme sensitivity to bending perturbations severely hinders practical applications. Starting from mode coupling theory, we theoretically prove and experimentally verify that higher-order modes exhibit significantly greater bending robustness than lower-order modes. We reveal that bending-induced changes in the output speckle arise predominantly from alterations in the mode effective propagation constants, whereas the modal power redistribution caused by inter-modal coupling is negligible. Since the bending-induced changes are dominated by alterations in the mode effective propagation constants, which are inherently independent of the input field, and the modal power redistribution caused by inter-modal coupling is negligible, it follows that, under moderate bending, the perturbation imposed by bending on light propagation inside the fiber is independent of the input field distribution. Based on this input-independent bending perturbation property, we propose a physics-inspired dual-encoder neural network. The network separately extracts features from a reference speckle pattern (corresponding to a circular field input) and an information-carrying speckle pattern acquired under the same fiber state, then employs a differential fusion module to decouple the image information from environmental perturbations, and subsequently recovers the image after excluding the perturbation. Our method significantly enhances the anti-bending capability of MMF imaging systems, offering a new paradigm that integrates physical insights with deep learning for robust fiber-optic imaging.