Research On Image Denoising Based On Deep Learning

Image noise reduction is a basic technology in image processing,and it is also a difficult problem in the image preprocessing stage.Image processing is widely used in medical images.Medical images are the basis and auxiliary tools for doctors to judge the lesion and diagnose the condition.Therefore,research on how to better remove medical image noise has important practical significance.On the basis of analyzing and researching the existing image noise reduction methods,it is found that the existing methods still have some shortcomings,such as texture blur after noise reduction,less research on multiple noise situations,and lack of application to real images.The thesis mainly studies how to use neural network to denoise images.The main content includes neural network construction,parameter setting,convolutional network denoising,data preprocessing,training and evaluation,etc.Including:(1)Constructing a convolutional autoencoding network with residual modules.The network consists of two parts,the encoder and the decoder.The encoder is composed of multiple residual modules with similar structures.Each residual module contains two branches: a residual learning branch and an identity mapping branch.The residual module is composed of a basic Re LU function and a pair of convolutional layers.Unlike the traditional residual module,this module uses the Re LU function before each convolutional layer,which helps the trained network to converge more easily and the resulting network More universal.Then use identity mapping as a jump connection,so that forward or backward signals can propagate directly between modules,which promotes the stability of training.(2)In order to reduce the loss of deep detail information and edge information in the input picture,the expansion convolution is used to construct a suitable convolution filter.Use multiple expansion rate convolutions to increase the receptive field of the convolutional layer,and the expansion rate of each expansion convolution is different,so that the neurons in the final layer can observe input features of different scales,reducing the loss of high-frequency details,thereby reducing The training error of the small network.The paper conducts an experimental comparative analysis on the public dental ray dataset(DX).The experimental results show that the training loss and verification loss are increased by 7.8% and 6.13%,respectively.The data is subjected to experimental comparative analysis under different degrees and different types of noise,and PSNR and SSIM are better than convolutional self-encoding networks;In addition,good visual effects have been achieved on the real dental images.The convolutional autoencoding network with residual module proposed in this paper is effective.