MT-SCUNet: A hybrid neural network for enhanced mode decomposition in optical fibers

Abstract

In few-mode and multimode fibers, data-driven mode decomposition (MD) leveraging deep learning has achieved notable progress and demonstrated significant advantages in simulated environments. However, when applied to experimental scenarios, the practicality of MD is hindered by substantial challenges, primarily due to inherent noise and alignment errors in optical profile image acquisition systems. Therefore, MT-SCUNet: a multitasking hybrid neural network model by integrating Swin-Transformer and Convolutional Neural Network architectures is proposed in this paper to address these limitations. It is capable of performing image restoration, classification and modal coefficient-based image reconstruction tasks simultaneously. Furthermore, accurate predictions on real-world images are attained upon the convergence of training, facilitated by the meticulous processing of pure simulation data with additive white Gaussian noise (AWGN) and mismatch errors to align with experimental conditions. The model’s effectiveness is verified using both simulation and experimental data on a few-mode fiber supporting up to 10 modes. The results show that the model performs well in terms of image restoration and reconstruction accuracy, with average peak signal-to-noise ratio (PSNR), structured similarity index measure (SSIM), and Pearson correlation coefficient (PCC) values of 66.03 dB, 0.9824, and 0.9910 for simulated data and 60.03 dB, 0.9694, and 0.9733 for experimental data, respectively. Additionally, the model also achieves 99.11% classification accuracy on the validation set. This work provides a solid foundation for advancing data-driven deep learning algorithms in MD, while also opening up new possibilities for applications in optical communications, sensing, and imaging systems.