Towards optimal multimode fiber imaging by leveraging input polarization and deep learning

• Published in: Optical fiber technology 2024-10, Vol.87, p.103896, Article 103896
• Main Authors: Maqbool, Jawaria, Hasan, Syed Talal, Cheema, M. Imran
• Format: Article
• Language: English
• Subjects: Deep learning; Imaging; Input polarization; Multimode fibers
• ISSN: 1068-5200
• DOI: 10.1016/j.yofte.2024.103896

Deep learning techniques provide a plausible route towards achieving practical imaging through multimode fibers. However, the results obtained by these methods are often influenced by various physical factors such as temperature, fiber length, external perturbations, and the polarization state of the input light. While previous studies have explored the impact of these factors on deep-learning-enabled multimode imaging, the effects of input polarization remain largely unexplored. Here, we experimentally demonstrate that the polarization state of light injected at the input of a multimode fiber significantly affects the fidelity of reconstructed images from speckle patterns. Certain polarization states produce high-quality images at fiber output, while some yield degraded results. To address this, we have developed a conditional generative adversarial network (CGAN) capable of regenerating images under various degrees of input light polarization. Our model stands out by achieving an SSIM score over 0.9 with a 50μm multimode fiber and demonstrating superior performance in both short training (1 h) and inference times (9.4 ms), unlike earlier research that primarily focused on multimode fibers larger than 50μm. Furthermore, our results demonstrate that the model can be trained to produce adequate imaging results for all input light polarization states, even with bends or twists in the fiber. We believe that our findings represent a significant step towards developing high-resolution and minimally invasive multimode fiber endoscopes. •Demonstrated input polarization impact on multimode fiber imaging.•Developed a high-performance CGAN model for image reconstruction•The CGAN model exhibited short training duration (1 h) and inference time (9.4 ms).•Achieved SSIM score exceeding 0.9, even for a thin 50 μm core fiber.•Achieved SSIM value above 0.8 for varying input polarizations and fiber positions.