Wavelength Independent Image Classification through a Multimode Fiber using Deep Neural Networks

Abstract

Deep Neural Networks (DNNs) have been increasingly implemented in different research fields or industrial applications. Large amounts of data are processed daily in order to extract useful information using machine learning techniques. Many research groups have shown impressive results on improving resolution in microscopy and quantitative phase retrieval by training DNNs on real datasets [1,2]. Recently, recovery and reconstruction of images after they have propagated through multimode optical fibers (MMFs) have also been achieved using DNNs [3,4]. When images propagate through MMFs they suffer severe scrambling because the information gets distributed among the different spatial modes that the fiber supports. Furthermore, since the fiber modes propagate with different velocities, the local information of the input decorrelates after a few millimeters along the MMF, thus resulting in the formation of a speckle pattern at the output. Recovery of information from such speckle patterns is of practical interest for integrating the MMFs for endoscopic applications in medicine or for signal recovery in telecommunications.