Enhanced angular misalignment resiliency in multimode fiber imaging leveraging deep learning

Multimode fiber (MMF) enables high-fidelity speckle imaging due to its densely loaded modes for carrying information, making it highly applicable in industrial inspection and medical diagnostics. Nevertheless, a slight misalignment in the angle of light incidence in the real environment directly affects the imaging quality, leading to changes in the excited mode number, orders, and coupling process. As the angular misalignment increases, the speckle pattern gradually becomes ring-like from circle-like, and features are mainly distributed at the edge areas. Existing works are typically designed for circular patterns where features are centered, and their performance is limited when facing cases with relatively large incidence angles. In this paper, to the best of our knowledge, we comprehensively investigate this phenomenon for the first time. We then propose a multi-residual Unet (MResUnet) deep-learning model to improve the imaging quality against the negative effects of increased incidence angular misalignment. The proposed scheme is realized by introducing multiple residuals so that the edge information is not completely masked by high-level features and more focusing on the relationship between the global and local speckle features. The results show that, compared to traditional methods, the accuracy of the test dataset is improved by 38% when the incident angle is even around 8°, and its structural similarity (SSIM) value reaches up to 0.96. All results indicate the great feasibility of the proposed MResUnet offering a steady way for high-quality MMF imaging.