Image Denoising Algorithm Based On Progressive Residual Dense Fusion And Quaternion Convolution Network

The image will inevitably be polluted by noise in the process of generation and transmission,which seriously affects the further analysis and processing of the image.Image denoising,as an important step in image processing algorithm,helps to further distinguish and interpret the image content,and is of great significance in the field of machine vision.In this paper,gaussian noise and random impulse noise removal algorithms are studied respectively.The main work is as follows:(1)Gaussian noise removal based on progressive residual fusion dense network:Denoising methods based on deep learning can achieve better denoising effects than traditional methods,but existing deep learning denoising methods often have the problem of too much computing complexity due to too deep network.To solve this problem,a progressive residual fusion dense network Dn RFD for gaussian noise removal is proposed.Firstly,dense blocks are used to learn the noise distribution in the image,and the network parameters are greatly reduced while the local features of the image are fully extracted.Then,by using the progressive strategy,the shallow convolutional features are connected with the deep feature short lines to form the residual fusion network,and more global features for noise are extracted.Finally,the output characteristic graph of each dense block is fused and input to the reconstruction output layer to obtain the final output result.The experimental results show that when the gaussian white noise level is 25 and 50,the network can obtain higher mean values of peak SNR and structural similarity,and the denoising average speed is twice higher than that of Dn CNN method and twice better than FFDNet method.In general,the denoising performance of this network is better than that of the correlation contrast algorithm,which can effectively remove gaussian white noise and natural noise in the image,and better restore the image edge and texture details.(2)Color random impulse noise removal based on two-channel quaternion convolution network: Deep learning-based color image denoising methods usually combine each channel after the convolution operation to get the final convolution result.This method does not fully consider the spectral correlation between color channels,which may lead to the distortion of denoising results.Quaternion convolution can solve this problem by treating color pixels as a whole.However,a single quaternion convolutional network can not restore image details well.To solve this problem,a two-channel quaternion convolution network(DQNet)is proposed to remove color random impulse noise.Firstly,based on the fusion strategy of structure channel and color channel,the structure details reduction module based on extended convolution is used to extract structure and edge features,and the quaternion convolution network is used to extract color features.Then,aiming at the problem that the convolution operation will cause partial global information loss,the long line connection is used to fuse the input noise image with the convolution result,and the feature enhancement module based on the attention mechanism is designed to guide the network to extract potential noise features in the complex background.Finally,residual learning is used to recover color random impulse noise.Experimental results show that the proposed algorithm has good denoising performance,and the denoising effect is more outstanding under the condition of moderate or high noise pollution.(3)Color random impulse noise removal based on quaternion convolution attention denoising network: Deep learning based color image denoising methods do not fully consider the spectral correlation between color channels,which may lead to a decrease in the restoration accuracy.To solve this problem,a quaternion convolution attentional denoising network(De QCANet)was proposed to remove color random impulse noise.In this network,a structure information extraction module based on extended convolution is designed to extract structure and detail features,which are taken as the real part and the three channels of color noise image as the imaginary part to construct a new quaternion matrix.Using this strategy,the quaternion convolution network can further obtain cross-channel color information.Finally,combining global and local features,a feature enhancement module based on attention mechanism is proposed to guide the network to remove color random pulse noise.Experimental results show that the new algorithm has good denoising performance.