Block Term Decomposition Based Hyperspectral Image Fusion And Denoising

As an emerging remote sensing imaging technique,hyperspectral imaging has received extensive attention due to its high spectral resolution.However,since hyperspectral imagers need to acquire photons from multiple spectral bands,its spatial resolution is usually low.In order to obtain super-resolution images with high spatial and spectral resolution,one common way is to fuse hyperspectral images with multi-spectral images.In addition,hyperspectral images are often inevitably contaminated by various noises during the acquisition.Therefore,hyperspectral denoising has received extensive attention.However,the traditional hyperspectral fusion and denoising methods firstly unfold the hyperspectral image into a two-dimensional matrix,and tackle it with traditional image processing methods.This operation would lose the inherent structural information within the hyperspectral image.Tensor,as a multidimensional extension of the matrix,has been proven to be more effective in many multi-dimensional data processing tasks.In this paper,based on the matrix-vector tensor factorization?MVTF?which is a special form of block term decomposition?BTD?,we have done research works on hyperspectral image processing as follows:1.Image fusion based on high spatial resolution multispectral images and low spatial resolution hyperspectral images.In order to obtain remote sensing images with both high spatial and high spectral resolution,we propose an image fusion method based on MVTF.This method employs MVTF for modeling the hyperspectral,multi-spectral and super-resolution images,and the image super-resolution problem can be transformed into a coupled MVTF problem.Compared with existing fusion algorithms based on tensor decomposition,MVTF can fully explore the internal structure of hyperspectral data,and better fit the linear mixed model of hyperspectral images.Meanwhile,in order to accelerate the proposed algorithm,the Sylvester equation is introduced to avoid iteratively updating a factor matrix.The experimental results show that the obtained super-resolution images are rich in detail,and have performance improvement in terms of RSNR and SAM,compared with existing methods.2.Hyperspectral denoising based on MVTF.For the noise-free hyperspectral image,we use MVTF based term to model the global correlation among all bands,and a total variation regularization to characterize the piecewise smooth structure in both spatial and spectral domains.Then the?₁ norm and Frobenius norm are used to constrain the sparse noise and Gaussian noise respectively to remove various noise.The well-known augmented Lagrange multiplier?ALM?method is used for solving the problem.Numerical experiments on both simulated and real-world data illustrate that the proposed method gives superior the HSI restoration results over some other popular ones.