The Research Of Multimode Fiber Imaging Based On Deep Learning

The research on multimode fiber mode transmission has a history of more than half a century.Even today,multimode fiber imaging is not mature enough to be commercialized compared to the single-mode fiber bundles that almost all endoscopes use every day,but at the same time,due to its high capacity,minimal invasiveness,and no footprint and other advantages,it is still attractive.In principle the transmission of a multimode fiber can be completely determined by a pixel-to-pixel transmission matrix connecting amplitude and phase(rather than intensity)from near-end to far-end,however,even with digital holography,it inevitably involves interferometry,the measurement and modulation of optical phase has never been an easy task.Artificial neural networks offer a unique way to achieve this.By recording the input image and output speckle at the near and far ends of the multimode fiber,respectively,the neural network can be trained using these pairs of modes to learn the corresponding pixel-to-pixel intensity maps between them that characterize the multimode fiber in a particular deployment configuration spread behavior.Once converged,the neural network can predict the output speckle of the new input image,enabling multimode fiber mode transmission.Transmitting natural scenes through multimode fibers is both beneficial and challenging for the fields of optics and artificial intelligence.Machine learning makes it possible to reconstruct patterns from intensityonly data,which avoids the interference of manual measurements.However,in order to overcome the interference characteristics of multimode fibers,low-cost and fast-training neural networks are urgently needed.This paper proposes a high-performance neural network model suitable for multimode fiber imaging,which has the advantages of high image transmission quality,fast training speed,and moderate storage overhead.The innovations and main work of the paper are as follows:1.Propose a high-performance neural network model for multimode fiber imagingIn the study,a well-designed hybrid neural network structure consisting of real-valued fully connected layers and shallow convolutional layers is proposed for multimode optical fiber natural scene transmission.Compared with other state-of-the-art research and similar schemes,this scheme is shown to have a faithful recovery ability to the most complex natural scene patterns,while significantly speeding up the training process.Considering the unique challenges of multimode fibers to artificial neural networks,the new design principles may help solve not only the problem of mode transmission in multimode fibers,but also the solution of similar problems in the wider field of optics and machine learning.2.Design a multimode fiber imaging simulation system to explore the influence of transmission matrix,amplitude and phase on imaging qualityThe real multimode fiber imaging system is unstable,and it is meaningful to analyze the influence of the reconstruction quality of the neural network model for multimode fiber imaging when the transmission matrix,intensity and phase are affected by different interference levels.In this study,by establishing a multimode fiber simulation system,the noise in different aspects of the system can be accurately simulated.On this basis,the robustness of the model is effectively analyzed and the generalization ability of the model is further proved.3.Through the network model comparison test and linear degree test,analyze the influence of the network module on the final imaging performanceThis study reproduces and compares typical models,explores the influence of different network modules on performance parameters,analyzes the role of each module in the high-performance model,and through the model linearity degree experiments to verify that although linearity plays an important role,the high-performance model Having more nonlinear properties can increase the upper limit of the restoration quality.This series of experiments improves the interpretability of the model and provides directions for how to construct neural network models when faced with similar problems such as imaging of other scattering media.