Research On Image Reconstruction Of Multimode Fiber Based On Deep Learning

Traditional target detection is limited by the entrance pupil diameter and the low-temperature refrigeration of the focal plane,making it difficult to achieve miniaturization and portability.Multimode fiber(MMF)can effectively solve this problem because of its flexible transmission and strong anti-interference characteristics.The deep learning method based on convolutional neural network(CNN)can effectively reconstruct the input image from the speckle image formed by the multimode fiber.Compared with the traditional methods such as transmission matrix,phase conjugation,and iterative algorithm,it is faster and more capable of overcoming complex external interference.In this paper,deep learning methods are used to carry out MMF image reconstruction research.First of all,aiming at the blank of the research on the modulation transfer function(MTF)of the MMF image transmission system,based on the sampling theory and the angular spectrum theory,the MTF of the MMF is deduced,and the MTF of the multimode fiber image transmission system is analyzed and the problem of image quality evaluation of image transmission system is solved.Secondly,according to this theory,a multimode fiber transmission imaging system in the visible light band is designed,and the establishment and debugging of the multimode optical fiber transmission imaging system is completed,and a sufficient number of speckle images are collected using this system as the data set of the neural network.Finally,for the image reconstruction of speckle images,based on the two convolutional neural networks of VGGNet and residual network(Res Net),two improved CNNS are designed and improved,and the effect of different parameter settings is discussed by comparing and studying the reconstructed image information,and finally the two improved CNNs are evaluated using PSNR and SSIM.The improved CNN based on the residual network and the VGGNet network can better complete the image reconstruction task,and the residual network improves the structural similarity of the reconstructed image to 88%,which is more suitable for the MMF image reconstruction task.