Research On Reconstruction And Imaging Of Fiber Speckle Based On Deep Learning

In recent years,Multimode Fiber(MMF)has been widely used in the fields of intraluminal medical diagnosis and optical imaging.Due to there are many modes in MMF and the interaction among them(such as interference and dispersion),the image presents a speckle-like pattern after being transmitted through the MMF.Therefore,clarifying the nonlinear mapping relationship between MMF images on the input and output is the key to study MMF image transfer and speckle reconstruction imaging.In this paper,a CNN network is designed to achieve high utilization of speckle information and high accuracy of speckle image reconstruction.In addition,many factors in the MMF imaging process will affect the amplitude and phase distribution of the speckle pattern,such as the length of the MMF,incident light intensity,etc.Therefore,in order to study the influence of different experimental conditions on data collection during the speckle collection process,we simulated the MMF transmission process and generated a numerical speckle data set through the multi-mode fusion method.The main research contents include the following two aspects:1.A multi-mode fusion method is proposed to simulate the MMF transmission process,and a numerical simulation speckle data set is generated.By analyzing the image correlation characteristics of the numerical speckle data set and the experimental speckle data set,it is proved that the numerical simulation speckle is consistent with the numerical statistics law of the speckle;Considering that U?Net is suitable for small datasets and its skip connections can better utilize the information in the image,the U?Net is used to realize the reconstruction experiment of numerically simulated speckle images.The experimental results show that U?Net can learn the nonlinear relationship between the numerical simulation speckle and the original label number,and realize the reconstruction of the numerical simulation speckle image.2.An improved deep neural network(AM?U?Net)is proposed,which adds the attention mechanism module to the down-sampling part of U?Net,and selects DSSIM as the loss function of the training process for the reconstruction of MMF output speckle images.By training AM?U?Net with only 300 label-speckle image pairs,it is proved that the model has excellent image reconstruction and generalization ability in the range of 1.2m?3.0m MMF length and 1.51 MW ? 18.3MW incident laser intensity.In addition,a bimodal fusion method is proposed to evaluate and improve the accuracy of image reconstruction,which combines the classification accuracy of both orthometric S-polarized and P-polarized speckle reconstructed images and forms a logical judgment.The accuracy of the speckle reconstructed image is increased to98.44%.These results all prove that the AM?U?Net has excellent optical speckle reconstruction ability and transfer learning ability,as well as good conditional adaptability to the transmission conditions of MMF.