A Variational Autoencoder for Bend-Robust Imaging through Multimode Fibers

Conventional endoscopes use fiber bundles, mapping each pixel to a separate fiber. Multimode fibers (MMFs) offer a compact alternative, encoding images via spatial modes within a single fiber. However, modal dispersion and intermodal coupling distort transmitted images into speckle patterns, further varying with fiber bending, complicating flexible imaging applications. We present a real-time imaging system using flexible MMFs, robust to bending, and requiring no distal-end access or feedback. A variational autoencoder (VAE) reconstructs and classifies images from speckle patterns, successfully generalizing to fiber bends unseen during training. Our system employs a 300 mm MMF (62.5 um core) to image 10 x 10 cm objects at a 20 cm distance, maintaining performance across fiber bends of up to 50 degrees and an 8 cm displacement range.