Hybrid vision transformer and convolution-based image reconstruction via multimode fiber speckle patterns of scaled Laguerre–Gaussian modes

In this study, we utilize size-scaled Laguerre–Gaussian (LG) modes to achieve high-fidelity image reconstruction through multimode fibers (MMFs), despite the severe intermodal coupling and phase scrambling that transform coherent images into complex speckle patterns. To perform this recovery, we employ a hybrid architecture combining vision transformers (ViT) and convolutional neural networks (CNNs), leveraging the CNN’s local feature extraction and the transformer’s global self-attention mechanisms. Using this framework, we demonstrate that high-fidelity image reconstruction with scaled LG modes performs significantly better than the standard Gaussian mode. Our results show that uniform coupling of the modes in the MMF achieved by size-scaled modes maintains higher and uniform coupling across different topological charges, resulting in a slight reduction in performance at higher orders while keeping both SSIM and PCC above 0.90. Also, we have compared our results with the corresponding conventional unscaled LG modes. These findings confirm that ViT-CNN models combined with size-scaled LG modal distributions provide a robust solution for image reconstruction via MMFs.