Learning to see high-density random images long-term transmitted in multimode fiber

Abstract

An improved multi-channel symmetric network (MCSNet) is proposed to reconstruct high-channel-density random images after long-term transmission through multimode fibers (MMFs). Temporal correlation within a period of 25 minutes is calculated to investigate the time-varying characteristics of speckles. The results demonstrated that due to noise accumulation along the MMF path, the quality of speckles deteriorates significantly after long-term transmission. The MCSNet integrates U-Net and ConvNeXt Block, which enables to more fully extract the features of each channel within the entire speckle. After being trained by different random image datasets within the initial moment, tests on random images and realistic scenes of endoscopic surgery after 25 min of transmission are carried out, and all of them demonstrate a near-perfect reconstruction performance and superior scalability, which indicates that MCSNet is suitable for long-term imaging demodulation of endoscopes.