Research On Image Denoising Algorithm Based On Deep Sparse Low Rank Network

Image denoising is an important research direction in image processing,which is widely used in medical diagnosis,security monitoring and remote sensing images.Image denoising algorithms mainly include model-based methods and deep learning methods.The model-based denoising method defines the image denoising model through the prior knowledge of the image,and the denoising model has an accurate mathematical definition.Based on the deep learning method,the noise image is directly mapped to the clean image through the deep neural network,which can achieve better performance improvement,but the interpretability of the denoising network is poor.In order to effectively integrate the advantages of these two methods,this paper uses algorithm expansion technology to study the denoising network guided by deep sparsity and low rank model.The main research contents are as follows:(1)An image denoising network based on deep double-layer group sparse coding is proposed.Firstly,a Two-Layer Group Sparse Coding(TLGSC)model is proposed based on the group sparse coding model.The core idea of this model is to further sparse code the group sparse coefficients to improve the representation ability of complex structures in images.Then,the iterative optimization steps of TLGSC model are expanded into two-layer group sparse coding network(TLGSC-Net)by using algorithm expansion technology.Through the analysis of synthetic noise experimental results,the image denoised by TLGSC-Net can retain more detailed information,and it has better denoising effect than the existing denoising algorithm based on algorithm expansion.(2)A hyperspectral image denoising network based on deep dynamic low rank matrix decomposition is proposed.Considering the low-rank characteristics of hyperspectral images and effectively representing the structural characteristics of images,a dynamic low-rank matrix factorization(DLRF)model is proposed,and then the iterative optimization solution of DLRF model is expanded into a dynamic low-rank matrix factorization network(DLRF-Net)by using algorithm expansion technology.In addition,the weight distribution branch is used to dynamically distribute the weight for DLRF-Net.By analyzing the experimental results of synthetic noise and real noise,DLRF-Net has certain denoising advantages compared with the contrast algorithm of hyperspectral image denoising.(3)A hyperspectral image denoising network based on deep low-rank sparse coding is proposed.Aiming at the lack of knowledge of hyperspectral image denoising algorithm based on deep learning,a Low Rank Sparse Coding model(LRSC)is proposed,and then the iterative optimization solution of LRSC model is expanded into a Low Rank Sparse Coding Network(LRSC-Net)by using algorithm expansion technology,in which the network reasoning process and iterative optimization process are the same,and the network parameters and optimization model parameters also have the same mathematical definitions,so the network parameters are interpretable.In addition,through the experimental results of synthetic noise and real noise,LRSC-Net has achieved better results than the contrast algorithm in detail preservation and clarity.