Research On Interpretable Deep Neurnal Networks For Image Denoising

In recent years,deep learning has made remarkable progress in image denoising,which has attracted extensive attention.However,most of the existing deep learning based image denoising methods are highly data-driven,and their parameters are not interpretable.In contrast,the traditional model-based denoising methods can define the model accurately and have the strong interpretability.A deep neural network driven by model and data can effectively integrate the advantages of the two methods,and it is an effective way to solve the interpretability problem of deep neural networks.Therefore,this paper uses deep unfolding idea to develop interpretable deep neural networks based on image denoising problems,which has important theoretical and practical application value.The specific research contents are as follows:1)A deep denoising network driven by group sparse coding model is proposed.Firstly,a regularized side group sparse coding model(Side Group Sparse Coding,SGSC)is proposed,which can effectively improve the representation of local and non-local structures in images by using auxiliary regularized terms constructed from pre-processed sparse coefficients.Then,an optimization method of alternating iterative algorithm is proposed to solve the SGSC model.Finally,the optimization solution of SGSC model is converted into a deep neural network(Side Group Sparse Coding Inspired Network,SGSC-Net)by deep unfolding strategy,in which each layer of the network corresponds to each step of the optimization solution.Therefore,the proposed SGSC-Net network parameters have accurate mathematical definition and can be interpreted.Experimental results on BSD68,Set12,CBSD68,Kodak24 and other public data sets show that the proposed SGSC-Net is superior to the existing deep unrolling based image denoising methods,and has certain competitiveness compared with the current popular deep neural network based on manual design.2)A deep denoising network based on multi-scale convolution sparse coding is proposed.Firstly,based on the original convolution sparse coding model,the multi-scale convolution sparse coding model(Multi-Scale Convolutional Sparse Coding,MSCSC)is constructed by introducing the multi-scale convolution dictionary,which can describe the multi-scale structure characteristics of images effectively.Then,the traditional iterative optimization solution of multi-scale convolutional sparse coding model is transformed into a deep neural network architecture,namely MSCSC-Net(Multi-Scale Convolutional Sparse Coding Network),by using the deep unfolding idea.Each layer in MSCSC-Net corresponds to each iteration of the optimized solution.In addition,in order to retain the structure information in the original image more effectively,MSCSC-Net adopts an improved residual learning idea to further improve the denoising effect of the network.Experimental results on BSD68,Set12,CBSD68,Kodak24 and other public data sets show that MSCSC-Net is superior to existing deep-unfurl image denoising methods.3)A deep denoising network based on 3D convolution sparse coding is proposed.Firstly,the original convolutional sparse coding model is extended to 3D form for multi-spectral image denoising,and 3D convolutional sparse coding denoising model(3D Convolutional Sparse Coding,3DCSC)is built.Then,the optimized solution is transformed into a 3D convolutional sparse coding network(3DCSC Network,3DCSC-Net).Experiments on CAVE dataset show that compared with CSCNet(Convolutional Sparse Coding Network),3DCSC-Net has better denoising performance.