Image Transmission Through Multimode Fibers By Deep Learning

Multimode fiber can deliver much higher information density than the single-mode fiber over the core diameter of tens of microns.The most prominent and valuable application of multimode fiber imaging is multimode fiber endoscopy,which can look deep into tissues including the brain.Compared with the traditional optical imaging technology,the flexible multi-mode fiber can flexibly change the optical path,therefore which could image at different positions.However,the multi-mode fiber can’t be directly imaged,and the problems such as dispersion and mode coupling in the transmission process leading to the disrupted speckle emitted from the far end of the multi-mode fiber.Currently,techniques and architectures for measuring transmission matrices,wavefront shaping and compression imaging have been proposed and experimented,but these methods are troubled by measuring complex optical fields,fiber disturbances and imaging efficiency.With the rapid development of deep learning technology in recent years,it provides a data-driven end-to-end multi-mode fiber image transmission scheme,avoiding complex optical measurement.However,the traditional deep learning method is still limited by the difficulty of data collection and slow training speed.Moreover,when the multi-mode fiber is deformed and disturbed,the image reconstruction quality of the deep learning model completed by training is seriously degraded.In this thesis,a model based on self-attention mechanism and hybrid training method are proposed to provide a high quality and anti-disturbance imaging scheme for the existing problems of multi-mode fiber image transmission.The main innovation points and main work of this thesis are as follows:A self-attentional model is proposed for high quality multimode fiber image transmission.In order to improve the reconstruction quality of multi-mode optical fiber image transmission,we introduce self-attention mechanism and design a neural network model based on multi-head selfattention.By introducing shifted window self-attention block into the model,the model can better learn the long-distance dependence in the sequence,so as to improve the quality of image reconstruction.In the multi-mode fiber transmission experiment,the SSIM and EME of image reconstruction quality on the transmitted datasets are increased by 0.04 and 0.79 respectively,and the number of parameters is reduced by 25%.The results show that the proposed method is one of the most efficient models at present,which can effectively reconstruct gray image and binary image,and is helpful to the solution of multi-mode fiber image transmission and other computational imaging problems.A hybrid training method is proposed to help the model resist multi-mode fiber deformation.Aiming at the robustness of neural network model to multi-mode fiber bending in image transmission,we make simulation dataset based on the analysis of experimental data,and use a large number of simulation data to prove that the hybrid training method can improve the anti-disturbance ability of the model in the multi-mode fiber transmission of high resolution image,and the reconstruction quality index SSIM of the data set is increased by 0.18 under different disturbances that do not participate in the model training.The method proposed in the thesis may pave the way for a simpler and more robust single multi-mode fiber image transmission scheme,which has the potential to be suitable for a variety of demanding image transmission tasks.