Fig. 1. Transmission characteristics of multimode fiber. (a) Multiple scattering effect of multimode fiber; (b) characterization of the input-output relationship of multimode fiber based on transmission matrix theory

Fig. 2. Light field reconstruction and projection in multimode fibers based on neural networks. (a) Classification and reconstruction of input light fields in multimode fibers by deep neural networks^［38］, where (i) shows the input and output light fields, (ii) indicates the VGG network used for classification and classification results, and (iii) represents the U-Net used for reconstruction and reconstruction results; (b) experimental setup for multimode fiber information transmission and projection^［39］, where (i) is the experimental setup, (ii) shows the projection results at the output end, and (iii) indicates the reconstruction results at the input end

Fig. 3. High-fidelity information transmission in multimode fibers based on deep neural networks. (a) Multimode fiber imaging combining principal component analysis and deep learning^[40], where (i) is the algorithm flowchart and (ii) shows reconstruction results and indexes; (b) high-fidelity image reconstruction via a multimode fiber inverse-scattering network^[41], where (i) presents the network structure and (ii) shows reconstruction results and indexes;(c) U-architecture speckle imaging network and its reconstruction outcomes^[42], where (i) details the network structure and (ii) visualizes the reconstruction results

Fig. 4. High-speed light field information transmission in multimode fibers based on deep learning. (a) Multimode fiber information transmission combining deep learning and transmission matrices^[44], where (i) is the experimental and algorithmic schematic, and (ii) shows inputs, outputs, and reconstruction results;(b) high-speed all-fiber image information detection system^[45], where (i) presents the data processing workflow and (ii) shows input images, output waveforms, and reconstruction results; (c) high-speed all-fiber imaging of small scenes with large depth of field^[46], where (i) is the system construction schematic, (ii) illustrates close-up imaging with large depth of field, and (iii) shows reconstruction capabilities at different depths of field

Fig. 5. Deep learning-based perturbation-resilient information transmission in multimode fibers. (a) Dynamic multimode fiber information transmission using deep learning^[47], where (i) is the dynamic multimode fiber schematic, and (ii) shows information reconstruction results for fibers of various shapes; (b) continuous multimode fiber information transmission based on semi-supervised learning^[48], where (i) shows the drift and variation characteristics of stationary multimode fibers, (ii) is the update process of the semi-supervised model, (iii) shows reconstruction results for speckles of different resolutions, and (iv) compares the performance of the semi-supervised and static models; (c) fast spatial pilot framework for long-distance stable multimode fiber transmission^[49], where (i) is the pilot fine-tuning schematic, and (ii) shows performance improvement after adding pilot fine-tuning; (d) long-term stable information transmission based on a multi-scale memory dynamic learning network^[50], where (i) is the time sequence update schematic, and (ii) shows the update process of the multi-scale memory dynamic network; (e) multimode fiber information transmission based on active disturbance measurement^[51], where (i) is the distribution schematic of single-mode fibers for monitoring, and (ii) is the experimental setup for image transmission and single-pixel imaging

Fig. 6. Multi-dimensional light field information transmission in multimode fibers based on deep learning.(a) Single-shot 3D phase information reconstruction using deep learning^[52], where (i) is the concept schematic, (ii) is the experimental setup, and (iii) is the 3D phase reconstruction results; (b) multimode fiber transmission of multi-spectral and multi-polarization channel information based on deep learning^[53], where (i) is the multi-spectral and multi-polarization schematic and (ii) is the related reconstruction results; (c) non-orthogonal optical information multiplexing transmission in multimode fibers via deep learning^[54], where (i) is the non-orthogonal concept schematic, (ii) is the speckle light field reconstruction network structure, and (iii) is the natural image reconstruction results

Fig. 7. Deep learning-based multimode fiber information transmission techniques enhanced by physical priors. (a) Bessel-equivariant network for achieving multimode fiber transmission inversion^[43], where (i) is the overall neural network architecture and (ii) shows visualized reconstruction results; (b) deep learning-based reflective imaging through multimode fibers^[4], where (i) is the framework for reflective imaging implementation and (ii) shows imaging effects; (c) deep empirical neural network integrated with transmission matrix^[55], where (i) is the overall framework of the deep empirical neural network and the accuracy evaluation of the empirical model and (ii) shows reconstruction results

Fig. 8. Performance optimization of multimode fiber information transmission. (a) Impact of light source properties on multimode fiber information transmission^[56], where (i) is the spatial resolution of different light sources and (ii) is the visualized reconstruction outcomes and assessment metrics; (b) efficient local speckle multimode fiber information transmission^[57]; (c) multimode fiber information transmission with transfer learning^[58], where (i) is the transfer learning schematic and (ii) is the reconstruction results of transfer learning with limited data; (d) improved perturbation resistance of broadband light sources^[59], where (i) is the experimental setup and (ii) compares speckles of LED and laser light sources; (e) all-optical image transmission in multimode fibers based on diffractive neural networks^[60], where (i) is the fiber-integrated optical diffractive neural network schematic and (ii) is the reconstruction results of all-optical demodulation