Super-resolution Of Hyperspectral Image Based On Tensor Graph Regularization

Hyperspectral image has high spectral resolution,however,its low spatial resolution affects the subsequent applications seriously.Multispectral image obtained in the same scene has higher spatial resolution and low spectral resolution,which is complementary to the hyperspectral image.By fusing the multispectral image and hyperspectral image,a more comprehensive and accurate understanding of the observed environment can be obtained.Therefore,it has important theoretical significance and application value to study the superresolution of hyperspectral image by the fusion of hyperspectral and multispectral images.The inherent three-dimensional structure of hyperspectral image can be effectively modeled by tensor.The key point to improve the super-resolution quality of hyperspectral image is to exploit the spatial-spectral prior by utilizing the interpretable meaning of factor matrices or tensors under tensor decomposition.This paper focuses on exploiting the prior of spatial factor matrices under Block Term decomposition of tensor.The graph Laplacian and super-pixel graph Laplacian priors are constructed by analyzing the correlation between hyperspectral and multispectral images in low dimensional manifold subspace.Moreover,new super-resolution models based on graph regularization of hyperspectral image are proposed.The main works of this paper are as follows:(1)The high resolution hyperspectral and multispectral images have the same spatial structure,which belong to a low dimensional manifold subspace.It is the most important issue to design the graph Laplacian to represent the manifold structure under tensor decomposition.Firstly,the ability of factor matrices obtained by Block Term decomposition in representing the spatial-spectral information of hyperspectral image is exploited.Then,the spatial similarity of multispectral image is utilized to represent the spatial structure of hyperspectral factor matrices by constructing the graph Laplacian prior,so as to make the "consistency" geometric structure between multispectral and hyperspectral images.Finally,the graph Laplacian regularization based coupled Block Term decomposition super-resolution model is proposed and solved by the block coordinate descent algorithm.Numerical results show that the proposed method has good performance in spatial detail promotion and spectral information preservation.Furthermore,analysis of the parameters involved in the experiment shows that the proposed method is robust.(2)In order to improve the preservation of the spatial local geometry structure in the fused image,the graph Laplacian regularization based on super-pixel is introduced into the super-resolution model.Firstly,the multispectral image is segmented by regional clustering,which is based on the spatial-spectral distance.The local geometric structure of super-pixel blocks is consistent with the hyperspectral image.Then,by extracting the horizontal and vertical feature tensors of the super-pixel segmented image,the spatial similarity weights of the super-pixel blocks are calculated and the graph Laplacian matrices are constructed.By conveying the spatial manifold information from the multispectral image to the spatial factor matrices of hyperspectral image,the spatial manifold structure between hyperspectral and multispectral images are well maintained.Finally,the proposed super-pixel graph Laplacian based Block Term decomposition super-resolution model is proposed and solved by the block coordinate descent algorithm.Compared with the model proposed in Chapter 3,the new model proposed in this chapter performs better fusion effect and stability by numerical experiment results,especially on the improvement of the spatial structure.