Research On Image Denoising Algorithm Based On Non-subsampled Contourlet Transform And Deep Learning

Image noise removal is a very important part of image processing.In today's world,with the rapid progress of science and technology and the rapid development of Internet,there are more and more image data we exposed to,and the desire for high quality images is increasing.It is very important to reduce the noise of the image because only high quality images can ensure the information complete in the image.At present,there are many methods for image de-noising.For traditional de-noising algorithms,there are various transform domain processing and de-noising algorithm based on statistical models,besides simple filter processing.In recent years,the emergence and development of deep learning has opened a new door for image processing.Therefore,deep learning based image de-noising algorithm has also emerged.This thesis mainly studies image de-noising from two aspects: One is processing the image in transform domain,and then using the modeling method of statistical model to realize image denoising;The other is to use depth learning to denoise the image.The work of this thesis is summarized as the following parts:(1)First,the image is processed by non-subsampled Contourlet transform,and the transform coefficient of the image is obtained,then the coefficient is modeled.By improving the estimation of the sub-band noise standard deviation in the Laplace modeling process,we also introduce the moderating factors for normal inverse Gauss modeling,so as to achieve the optimization effect of the model.(2)After modeling the image in the transform domain and modeling the normal inverse Gauss,the obtained modeling data are processed jointly by certain rules to get the joint modeling coefficient.Then the coefficient data obtained by the joint modeling are Inverse non-subsampled Contourlet transformation,and then the image after the denoising is obtained.(3)A double stage convolution neural network structure is proposed.The first stage of the network mainly completes the residual training of the image,that is,to get the noise of the image.Then we process the first stage image data through a certain operation,and then train it into the second stage network as input,and finally get the image de-noising effect diagram.(4)In double order convolutional neural networks,residual training and batch regular processing of data are not only retained,but also activation functions are modified.Here,the ELUs is used as an activation function instead of the Re LU function,so that the training speed of convolutional network in training process can be improved,and it can get better robustness.At the same time,in order to prevent the overfitting of the network,optimize the efficiency of the network training,and modify the loss function.