Small-dataset speckle feature transfer imaging in a multimode fiber

A Swin-ReconGAN network model is proposed to address the instability in MMF imaging caused by environmental disturbances. Speckle feature transfer is performed to adapt the pretrained reconstruction network, mitigating the impacts of fiber perturbations, speckle drift, and varying scattering media on image reconstruction. The proposed network achieves effective feature transfer imaging for individual fiber bending states using only 200 image-speckle pairs, significantly reducing the required training dataset size and data acquisition costs compared to traditional neural networks. After 50 independent runs in the cross-bending states feature transfer imaging, the Swin-ReconGAN achieved an average structural similarity index measure (SSIM) of 0.705, outperforming both the Transfer Learning U-Net (0.565) and the Scratch U-Net (0.620). Similarly, in the cross-medium feature transfer imaging, the Swin-ReconGAN achieved an average SSIM of 0.680 with a coefficient of variation (CV) of only 1.27%, significantly outperforming the Transfer Learning U-Net (average SSIM 0.497, CV 1.46%) and the Scratch U-Net (average SSIM 0.573, CV 9.84%). Beyond individual states, the Swin-ReconGAN also demonstrates robust capability in multiple discrete bending states feature transfer imaging. By directly performing feature transfer on speckle patterns, this model provides a practical approach for robust speckle reconstruction in small-sample scenarios.