Research On Fiber Speckle Image Restoration Based On U-net

In our life,the light we observe travels in straight lines.When light penetrates biological tissues,smoke,dust,clouds and other strong scattering media,it will scatter.The scattering phenomena will affect the clarity of light imaging.Because the scattered medium causes the light to scatter,the acquired image will be blurred,and the image will be degraded by noise.Therefore,it is very important to solve these factors to make the image clear.In the medical field,clear images play a vital role in the doctor's judgment.The common material of medical endoscope is optical fiber,and optical fiber,as a kind of medium,will degrade the image of the output end of medical endoscope.Therefore,studying the restoration of optical fiber speckle can not only improve the image quality of medical endoscope,but also greatly improve the efficiency and accuracy in the medical field.Because of the scattering effect of optical fiber,the image will be blurred.In medical endoscopy,the quality of the image will be affected and doctors' judgment will be disturbed.This article focuses on the restoration of optical fiber speckle images.First,it introduces the existing research methods of optical fiber speckle restoration,which can be roughly divided into traditional imaging technology and computational imaging technology.In traditional imaging technology,it is necessary to manually adjust optical experimental equipment to obtain a clear image.The process is extremely complicated and a large number of formula derivations are designed,which is time-consuming and labor-intensive.The computational imaging technology uses the powerful computing power of a computer and advanced image processing technology to process the obtained speckle images,which can obtain clear images more quickly and conveniently.Then,this article elaborates on the U-Net algorithm and its optimization algorithm of the applied full convolutional neural network in detail.As we all know,each layer of a neural network contains a large number of neurons,and the neural network may produce gradient explosion or gradient dispersion as the number of convolutional network layers increases.Therefore,in order to prevent the neural network from generating gradient explosion or gradient dispersion,the Residual Block structure is introduced here for optimization.At the same time,the quality of the restoration result of the speckle image lies in the use of each layer of the feature layer.In order to maximize the use of the features,the Dense Block structure is introduced for optimization.Finally,the basic principles of this experiment are introduced.From the perspective of physical experiments,optical instruments are used to build the experimental platform,set the corresponding experimental parameters,and take the corresponding optical fiber speckle to lay the foundation for the further experiment.Compare the network layer parameters,activation function,and loss function of the three neural network algorithms through experimental data to optimize settings,and carry the most suitable network architecture for the data.Input the preprocessed data into three kinds of neural networks for training,and finally realize restoration.By comparing the obtained restoration results,Dense U-Net has the best restoration effect,with an accuracy of 0.982 for the training set and 0.975 for the verification set.