Enhanced angular misalignment resiliency in multimode fiber imaging leveraging deep learning.

Area of Science:

• Optics and Photonics
• Machine Learning
• Image Processing

Background:

• Multimode fiber (MMF) is crucial for high-fidelity speckle imaging in industrial and medical fields.
• Angular misalignment of incident light significantly degrades MMF imaging quality, altering speckle patterns and feature distribution.
• Existing methods struggle with large incidence angles due to their focus on centered features.

Purpose of the Study:

• To comprehensively investigate the impact of incidence angular misalignment on MMF speckle imaging.
• To propose and validate a deep-learning model for improving MMF imaging quality under angular misalignment.
• To address the limitations of current methods in handling edge-distributed features caused by misalignment.

Main Methods:

• Investigated the phenomenon of speckle pattern changes with increasing angular misalignment in MMF.
• Developed a multi-residual Unet (MResUnet) deep-learning model incorporating multiple residuals.
• Focused on preserving edge information and the relationship between global and local speckle features within the MResUnet architecture.

Main Results:

• The MResUnet model demonstrated significant improvement in imaging quality against angular misalignment.
• Achieved a 38% accuracy improvement on the test dataset at an 8° incident angle compared to traditional methods.
• Reached a structural similarity (SSIM) value of up to 0.96, indicating high image fidelity.

Conclusions:

• The proposed MResUnet offers a robust solution for high-quality MMF imaging despite angular misalignment.
• The model's ability to handle edge features and global-local relationships is key to its effectiveness.
• Results confirm the feasibility and potential of MResUnet for practical industrial inspection and medical diagnostics.