| Hyperspectral imaging is a new analytical technique based on spectroscopy.It collects hundreds of images of different wavelengths for the same space region to collect the spatial and material attributes of the target space region.Therefore,hyperspectral images usually contain rich spectral information.Now,hyperspectral image has been widely used in environmental monitoring,biomedicine,oceanography,agriculture and other fields.However,due to the limitations of hardware equipment,the acquired hyperspectral images are usually have low spatial resolution,which makes it difficult for hyperspectral images to meet the requirements of detection,classification and other applications.Therefore,the problem of hyperspectral image super-resolution has become a hot topic in remote sensing and computer vision in recent years.The main idea of hyperspectral image super-resolution is to fuse low spatial resolution hyperspectral image with corresponding high spatial resolution multispectral image to obtain super-resolution hyperspectral image.Among them,the traditional hyperspectral image superresolution methods all need to unfold the three-dimensional hyperspectral image into matrix,and then analyze and process the data in the matrix form.However,these approaches destroy the spatial structure of the hyperspectral image itself,leading to the failure of the model to make full use of the correlation between spatial and spectral information in hyperspectral image.Therefore,approaches based on tensor decomposition is proposed to make up for the shortcoming of matrix decomposition in solving the problem of hyperspectral image super-resolution.In recent years,tensor analysis has been found to be an effective way for hyperspectral image processing,and hyperspectral image super-resolution methods based on tensor representation have also attracted more and more attention.Classical tensor decomposition models such as the CP tensor decomposition model and the Tucker decomposition model have been applied in this field.Inspired by this,two hyperspectral image super-resolution algorithms based on tensor ring decomposition are proposed in this paper.One is Coupled Tensor Ring Factorization based Hyperspectral Image Super-Resolution algorithm(CTRFHSR)and the another one is High-Order Tensor Ring Representation based Hyperspectral Image Super-Resolution algorithm(HOTRHSR).Due to the different low-rank properties of the three dimensions of hyperspectral images,previous proposed coupled matrix /CP decomposition based super-resolution models were unable to measure the low-rank properties of each mode of hyperspectral image,so these model can not accurately fit the data.To solve this problem,we proposed the tensor ring decomposition based model.The proposed CTRFHSR algorithm can not only measure the low rank of different dimensions in hyperspectral images by adjusting the rank of the tensor,but also has a more flexible representation compared with the algorithm based on coupled Tucker decomposition.Specifically,CTRFHSR assumes that super-resolution hyperspectral images can be represented as three core tensors by tensor ring decomposition and potential core tensors between the hyperspectral image and the multispectral image tensor ring decomposition are shared,and we use the relationship between the core tensors to reconstruct the super-resolution hyperspectral image.In addition,existing tensor representation based methods are unable to capture the high-order correlation of hyperspectral images.The HOTRHSR algorithm proposed in this paper can not only maintain the three-dimensional spatial structure of the hyperspectral image itself,but also explore the correlation between spatial information and spectral information in the hyperspectral image in the higher-order case.Moreover,for the above two models,we use the alternative optimization method to optimize the model.Finally,to test the validity of CTRFHSR and HOTRHSR algorithms,multiple datasets are used in the experiments.The performance of CTRFHSR and HOTRHSR is proved to be effective by comparing with the state-of-the-art algorithms. |
