Dynamic MR Image Reconstruction With Low-rank Tensor Constraints And Iterative Feature Refinement With Network-driven Prior For Image Restoration

As a fundamental problem in the field of data analysis and processing,the representation of data is extensively studied.Real data is constrained by multi-dimensional factors and has complex internal structure.Considering the high-latitude nature of these data,vector and matrix,which are commonly used in data representation,cannot well describe the global correlation of data and lose a lot of redundant information.Therefore,the application of multi-dimensional data has a wide range of research value.In this thesis,we apply high-dimensional images for image restoration(IR),and explore the validity of high-dimensional information from two directions:traditional model algorithms and deep learning algorithms.In the traditional model,the dynamic magnetic resonance imaging(MRI)problem is tackled by optimizing the algorithm and mining the prior information of the intrinsic structure of high-dimensional images.In the aspect of deep learning,high-dimensional images are utilized training the network and applied addressing the problem of natural image restoration.The researches are as follows:(1)Dynamic MR image reconstruction with low-rank tensor constraints:Tensors are introduced to describe high-dimensional data,and the low-rank tensor coding model is proposed for exploring the sparsity of MR images.Developing the local self-similarity of high-dimensional images,similar cubes are extracted and grouped constituting low-rank tensors.Algorithms such as augmented Lagrange multiplier(AL)and alternating direction multiplier method(ADMM)are introduced to address the model.This model can capture the sparse part of dynamic MR images,make full use of the redundant information between adjacent location feature vectors.The quality of reconstructed images is superior to the classic low-rank and sparse decomposition algorithm.(2)Iterative feature refinement with network-driven prior for image restoration:an unsupervised iterative feature refinement(IFR)model based on enhanced high-dimensional depth mean shift prior(EDMSP)is proposed.The model inherits the excellent noise suppression characteristics of embedded network and the fine detail preserving ability of IFR model.In the network training module,the R,G,B images are trained as samples to mine the high-dimensional information of the images and strengthen the generalization ability of the network.We apply the model to a variety of IR tasks,and the experimental results verify the robustness of the model.In summary,mining and using high-dimensional information can improve the performance of the model to a certain extent.With the development of related research,the development of high-dimensional information has great research value.