Research On Single Fiber Imaging Technology Based On Neural Network

Traditional endoscope technology uses a single-mode fiber bundle or a single single-mode fiber scanning method for imaging,but this type of endoscope has a large probe size,low resolution,and inflexible operation.Multimode fiber can accommodate the advantages of a large number of guided modes simultaneously transmitted in the fiber,so that a single multimode fiber can comple te imaging tasks independently.The endoscope using a single multi-mode optical fiber imaging can effectively overcome the shortcomings of the traditional endoscope,and realize a high-resolution ultra-fine endoscope.However,there are effects such as modal dispersion and mode coupling in the transmission process of multimode fiber,which will cause inter-mode interference in the optical field of different modes of the fiber,and the image transmitted in the fiber will be distorted or even become unrecognizable speckle.This article explores how to restore the distorted speckles at the output end of the multimode fiber to the original image at the input end to achieve single fiber image transmission.A method of fitting a multimode fiber transmission matrix with a neural network is proposed,and the input original light field is calculated through the reverse transmission matrix and the output speckle light field.This paper builds a multi-mode optical fiber imaging system,and uses the output speckle-input original image pair of the system to make a training set to train the neural network,so that the neural network fits the reverse transmission characteristics of the multi-mode optical fiber.The traditional real number network has limited fitting effect o n the multimode fiber transmission matrix.Inspired by the physical process of multimode fiber modulation of the light field,this paper also proposes to extend the network to the field of complex numbers,and realize complex number operations in the form of real numbers in the network.The neural network reproduces the same complex calculation as the physical process of multi-mode fiber image transmission,and the feasibility of the multi-mode fiber imaging method based on real-valued and complex-valued neural networks is verified by designing experiments for simple structure images and complex structure images.Aiming at the situation that the transmission matrix of multimode fiber changes with the shape of the fiber,the effectiveness of the transmission matrix based on neural network fitting under different degrees of fiber deformation is analyzed through experiments.Finally,this paper also studies the effective information distribution of the speckle emitted by the multimode fiber.Through the comparat ive analysis of the imaging results of the local speckles of the speckle in different regions,it is proved that the effective information that can be used to restore the original image is evenly distributed throughout the whole speckle.