Active Illuminated Multimode Fiber Imaging Based On Deep Learning

Optical fiber imaging is an optical imaging technology with small diameter,low loss and low cost.It is usually used for medical endoscopy through the human body natural tubes or surgical incisions,and can also be used for deep tissue imaging.In traditional optical fiber imaging technology,optical fiber bundles are usually used to transmit images.Each fiber core transmits information of one pixel,and the resolution of imaging system is limited by the diameter of the fiber bundles.In addition,traditional fiber endoscopes usually contain both illumination fiber and imaging fiber,which further increases the diameter of the transmission system.Multimode fiber has multiple modes and small core diameter,so it has been widely used in the field of optical fiber imaging technology in recent years.Multimode fiber can transmit a large number of guided modes simultaneously,so it can obtain higher imaging density than traditional fiber bundle imaging.However,as a result of the physical characteristics of multi-mode fiber,different guided modes will cause the loss of the original phase relation among the image information,and finally form a seemingly random spatial speckle distribution at the output end.In the task of original image reconstruction from speckle patterns,deep learning technique has been favored by many researchers for its faster reconstruction speed and better robustness.In this paper,an active illuminated multimode fiber imaging system based on deep learning technology is proposed.Firstly,a novel active illuminated multimode fiber imaging system is designed,which is more miniaturized compared with traditional external illumination multimode fiber imaging system.The system uses a single multimode fiber to implement illumination and image transmission.However,active illuminated system also brings more inherent noise,which increases the difficulty of image reconstruction.Therefore,a full-connected layerencoder-decoder network based on deep learning technology is designed for original image reconstruction from speckle patterns.This algorithm has stronger feature extraction and combination ability and is more suitable for active illuminated multimode fiber imaging system.In this paper,experiments on 1.6m long multimode fiber are conducted.We obtain shorter training time and better reconstruction results compared with the previous deep learning networks on three different datasets.The average training time of the proposed deep learning network is reduced by 63.90%,the average PSNR of reconstruction results is increased by 2.62dB,and the average SSIM is increased by 14.19%In addition,image reconstruction using partially cropped speckle patterns and the cross-domain generalization capability of the proposed deep learning network are also studied.All the above experimental results demonstrate the superiority of the system and the potential of the system for more miniaturized endoscopic imaging.