Adaptive Dynamic Learning for Efficient Long‐Term Image Transmission Through Unstable Multi‐Mode Fibers

Abstract

Multi‐mode fibers (MMFs) support high‐capacity optical transmission, making them ideal carriers for remote imaging and high‐throughput optical communication. However, the variable transmission channels of MMF pose major challenges for long‐distance, high‐precision spatial information delivery. Existing static modeling approaches fail to adapt to fiber variations over long time and suffer from severe performance degradation. Frequently updating transformation models could achieve dynamic tracking of the fiber states at the expense of a massive computational budget. In this work, we propose a self‐supervised adaptive dynamic learning (ADL) strategy for highly efficient and drift‐robust image transmission through unstable MMFs. By employing a lightweight physics‐informed discriminator to quantitatively monitor transmission variations, ADL triggers model updates only when necessary. We experimentally demonstrated image transmission over 1 h through a 1 km MMF without special stabilization, and achieved reconstruction accuracy over 99.8%. Crucially, ADL reduces computation time by 93.6% compared to fixed‐frequency dynamic learning strategies, without sacrificing performance. The proposed highly computationally efficient strategy establishes a stable paradigm for long‐term spatial information transmission through dynamic scattering media. Code is available on GitHub ( https://github.com/SII‐WZ/ADL ).