Small-dataset speckle feature transfer imaging in a multimode fiber.

Area of Science:

• Optics and Photonics
• Machine Learning
• Image Reconstruction

Background:

• Multi-mode fiber (MMF) imaging is susceptible to environmental disturbances, leading to instability.
• Traditional methods require extensive datasets and are sensitive to fiber perturbations, speckle drift, and scattering media variations.

Purpose of the Study:

• To develop a robust network model for stable MMF imaging.
• To mitigate the impact of environmental disturbances on image reconstruction.
• To reduce training dataset size and data acquisition costs.

Main Methods:

• Proposed a Swin-ReconGAN network model incorporating speckle feature transfer.
• Adapted a pretrained reconstruction network to handle environmental impacts.
• Utilized a small dataset (200 image-speckle pairs) for effective feature transfer imaging.

Main Results:

• Swin-ReconGAN achieved an average SSIM of 0.705 in cross-bending states, outperforming Transfer Learning U-Net (0.565) and Scratch U-Net (0.620).
• In cross-medium imaging, Swin-ReconGAN achieved an average SSIM of 0.680 with low CV (1.27%), outperforming alternatives.
• Demonstrated robust capability in multiple discrete bending states feature transfer imaging.

Conclusions:

• The Swin-ReconGAN model offers a practical approach for robust speckle reconstruction in small-sample scenarios.
• Speckle feature transfer effectively addresses MMF imaging instability caused by environmental factors.
• The model significantly reduces data requirements and improves reconstruction performance compared to existing methods.