Deep learning-based multimode fiber imaging in multispectral and multipolarimetric channels

Abstract

Multimode optical fiber (MMF) imaging is an emerging fiber imaging technology that has been developed during the last decade. In this work, we demonstrate deep-learning-based MMF imaging for multispectral and multipolarimetric channels. Specifically, by controlling the wavelength and polarization of the incident light of MMF, different spectral and polarization channels are constructed. We conduct MMF transmissive imaging experiments and record a large number of object-speckle pairs in each channel for neural network training. A neural network is trained to simultaneously reconstruct the intensity and classify the channel of objects in eight spectral or nine polarimetric channels. The average structural similarity (SSIM) of the image reconstruction in each spectral and polarimetric channel exceeded 0.9 with the accuracy of the channel classification exceeding 99.9%. By superimposing speckle patterns of different spectral channels, a new dataset is constructed for training, and the reconstruction of images containing multiwavelength information is also tested in eight spectral channels. Our findings have the potential to extend the application of MMF imaging with spectral and polarimetric information