Research On Residual Network Image Denoising Method With Dual-domain Information And Self-paced Learning

As a carrier for people to obtain information from the outside world,images contain richer information than texts.However,due to problems such as scanning and transmission equipment,the received images are often of poor quality,which seriously affect subsequent understand,analysis of images and other tasks.Therefore,image denoising technology has become an important topic in the field of computer vision.At the same time,this direction has also attracted the attention of many scholars,and using different technologies,many denoising algorithms have been proposed.Recently,convolutional neural networks have gained good application in the field of image denoising with their good feature learning ability,however,these denoising methods have some problems.For example,the loss of image detail information after denoising,these denoising methods perform better on artificially generated Gaussian noise images but perform poorly on images with different noise levels.In view of the above problems,this thesis studies the image denoising method,which is as follows:1.Residual network research using dual-domain information for interactive learning.Aiming at the problem that the pooling operation in convolutional neural networks easily uses the texture and detail high-frequency information in images as noise cancellation,this thesis proposes a dual-domain denoising Residual network(Dual-domain Denoising Convolutional Neural Network with Residual Learning,DDCNN).The model is mainly composed of two convolutional layers,and there is a jump connection between the two convolutional layers,which does not require batch normalization operations,so that the input image information is directly transmitted to the hidden layer,which reduces the path length of gradient transmission and alleviates the problem of gradient disappearance.2.Research on image blind denoising method integrating self-paced learning mechanism.Although the deep convolutional neural network has achieved a good denoising effect on additive Gaussian white noise images,these methods have limited model denoising performance due to overfitting of the model to synthesize noise when facing images with different noise levels.Therefore,in combination with the self-paced learning mechanism,this thesis proposes a model(Dual-domain Self-paced learning Denoising Convolutional Neural Network,DSDCNN)with the function of estimating noise for image blind denoising.The model includes a noise estimation subsystem based on the self-learning strategy and an improved non-blind denoiser network,and takes the noise level evaluation as a regularization term in the objective function of the non-blind denoiser network.Through joint optimization,the model can take into account both denoising effect and noise evaluation.At different noise levels,the model has a certain robustness,and the denoising effect can be achieved better than the model trained at a single noise level.3.In order to further improve the training efficiency of blind denoising network,a multi-scale feature extraction method combined with dilated convolution is proposed.For the traditional denoising method,due to perceive more contextual information,at the expense of the shrinkage of the receiving field,it is easy to increase the number of parameters,thereby reducing the efficiency of network training.In this thesis,multiple expansion convolutions of different expansion rates are used to extract features in parallel,increasing the acceptance field of the convolutional kernel without changing the parameter quantity,while keeping the size of the output feature mapping unchanged.Aiming at the complex structures in images with different noise levels,the multi-dilated block of multi-scale dilated convolutional feature extraction module(Multi-dilated block,Dblock)is proposed.It uses multi-dilations rate dilated convolutional fusion to fuse multi-scale information in the image,so as to strengthen the network to extract image detail information,avoid excessive parameterization problems caused by the deep network,thereby reducing the overall iteration of the denoising model and further improving the denoising efficiency of the model.