Research On Image Denoising Method Based On Convolutional Neural Network

With the advent of the third wave of Internet,image restoration,a kind of low vision task,has been introduced into the field of deep learning.Being as the preprocessing step of the high-level visual tasks such as image classification,image segmentation,image restoration is of great significance.Image restoration technology includes image denoising,image deblurring,image high-resolution reconstruction and so on.As the most basic image restoration technology,image denoising has strong theoretical and application value.Due to the influence of air pressure,humidity,light and other factors,there may be noise pollution in the photos taken by electronic equipment,which will damage the image quality.Image denoising technology can make the object far away from the influence of noise to a large extent,and ensure the accuracy and effectiveness of the information transmitted by the photo.In recent years,great progress has been made in image denoising vision task based on convolutional neural network.The success of existing image denoising methods mainly comes from the effective use of non-local module(NLM)and residual block(Res B).NLM can capture the global information of the image and effectively improve the expression ability of the network;And Res B can ensure the stable training of the network.Based on this,the main research work of this thesis is as followsFirstly,a multi-scale denoising method based on channel attention is proposed in this thesis.In this structure,multi-level down sampling encoder is used,and then multi-level up sampling decoder is used in turn.The residual block stacking technology is used for auxiliary feature extraction,and the channel attention mechanism and residual block are combined to adaptively obtain the deep favorable features of the image.The performance of network denoising is further improved.Secondly,a non-local attention network(NLAN)is proposed to achieve high quality image denoising.Firstly,NLM is combined with a convolution-deconvolution structure,which can capture more deep feature relationships without increasing the amount of computation,and further improve the expression ability of the network.Then,this thesis proposes a new spatial attention mechanism,which adaptively adjusts each pixel in the feature image by using the row information and column information in the feature image at the same time,and introduces it into the residual block,so as to improve the ability of network differential learning.Through a large number of experimental results,it is proved that the proposed method can not only remove the noise in the image thoroughly,but also protect the edges and details of the image.The experimental results show that the subjective visual and objective performance are greatly improved.