High Accuracy Transmission and Recognition of Complex Images through Multimode Fibers Using Deep Learning

Multimode fiber shows tremendous potential in promoting the microminiaturization of optical endoscopes. However, multimode transmission is quite sensitive to fiber deformations and environmental changes. High-accuracy transmission of complex images through a multimode fiber using traditional methods remains challenging research. Deep learning, which shows enormous vitality in optical imaging, may break through this limitation. Here, a deep neural network: U-architecture speckles imaging network (USINET) is presented to realize high accuracy reconstruction of complex images under different multimode fiber transmission conditions. Furthermore, a shallow neural network: convolutional neural network (CNN)-architecture speckles recognition network (CSRNET) is designed to realize high accuracy recognition for multiple categories of speckles at the output of multimode fiber under different bending states. The experimental results demonstrate that the proposed networks can realize high accuracy transmission and recognition of complex images through multimode fibers, which indicates the application prospect of multimode fibers combined with deep learning in minimally invasive medicine.

Automated summary

Multimode fiber has great potential in miniaturizing optical endoscopes, but it is sensitive to fiber deformations and environmental changes. High accuracy transmission and recognition of complex images through multimode fibers using traditional methods are challenging. This study presents two neural networks, USINET and CSRNET, for high accuracy reconstruction and recognition of complex images under different multimode fiber transmission conditions. The experimental results demonstrate the application prospect of multimode fibers combined with deep learning in minimally invasive medicine.