Rotational Memory Effect‐Inspired Radon Domain Learning Empowers Image Transmission through Multimode Fibers

Abstract Image transmission through multimode fibers (MMFs) poses a complex inversion challenge due to the intricate light transport and potential information loss. While recent advances in deep neural networks (DNNs) have shown promise in modeling the MMF input‐output relationship, commonly‐used networks such as fully connected (FC) models and convolutional neural networks (CNNs) fall short of leveraging the inherent sparsity of MMF systems, which leads to learning inefficiency and poor angular generalization. Here, an ultra‐compact Radon transform‐facilitated cross‐domain learning framework, called Radon Transmission network (RTMnet), is presented. Inspired by the physical sparsity in MMF's rotational memory effect, the Radon transform is applied to the captured speckle and performs physics‐guided learning of MMF image transmission in the sinogram domain. RTMnet enables high‐fidelity MMF image transmission with an order‐of‐magnitude reduction in computational demand compared to traditional DNN models. Arbitrarily rotated handwritten digit images can be faithfully reconstructed using a limited training data of only 7000 non‐rotated digits. This enhancement in learning efficiency underscores RTMnet's physics consistency and its potential to effectively generalize in resource‐constrained fiber‐based applications such as miniaturized endoscopy systems.