| Hyperspectral images contain hundreds of spectral bands,ranging from visible to infrared wavelengths.Hyperspectral images are rich in spectral information and are used in many common applications,such as classification,target detection,monitoring of the earth’s environment,disaster prediction and agricultural production.However,due to the limitation of hardware technology,it is difficult to obtain hyperspectral images with sufficient spatial information and spectral information.A feasible solution is to fuse images with images of low spatial information and sufficient spectral information with images of sufficient spatial information and multispectral information.An important criterion for hyperspectral image fusion is that the generated target image has the same spatial information as the high space multispectral image and has the same spectral information as the low space hyperspectral image.Based on this criterion,we propose a research method of hyperspectral image fusion based on traditional tensor decomposition and low-rank tensor depth prior.The method is verified by multiple data sets,and it is found that the method achieves good results in multiple data sets.The main research work of this paper is as follows:Based on tensor ring decomposition method,hyperspectral image can be decomposed into multiple core tensor factors.By applying low-rank regularization to multiple core tensor factors,each core tensor factor can be restricted.Finally,the fusion of target image can be realized by recombining multiple core tensor factors.Therefore,a fusion algorithm based on low-rank regularization of tensor ring model is proposed in this paper.By combining the tensor ring decomposition model and the regularization of low-rank tensors,we built the corresponding optimization model to estimate the corresponding core tensor factors.Among them,the spectral response function applied by the method was estimated by the corresponding algorithm.Through experimental results,we found that the fusion method proved its effectiveness in multiple data sets.Based on the function of tensor decomposition of fully connected networks,this paper further improves the limitations of the tensor ring model,and then proposes a fusion algorithm based on the tensor decomposition model of fully connected networks.We apply a new tensor decomposition method to expand our image,which can capture both the depth information and the high order information of the image.Firstly,we carry out the high order image,transform the image into a high order tensor,and then apply the tensor decomposition method of the fully connected network to the high order tensor,which can make use of the multi-scale information,high order and depth information of the image.At the same time,we also pay attention to the relation between adjacent higher-order core tensor factors and the spatial information of images.In order to maintain spectral information at the same time,we use Laplacian graph regularization to maintain our spectral information,so as to achieve better fusion effect.The experimental results show that we find the fusion method has a great improvement in the effect of the three data sets.Based on the rapid development of tensor decomposition in deep learning embedding,this paper proposes a deep fusion algorithm based on tensor low-rank prior knowledge.By embedding Canonical Polyadic(CP)into our matrix decomposition model,we design an optimization model combining traditional model and deep learning.In order to estimate the coefficient matrix,the model can be iterated through the deep learning network,and the number of iterations corresponds to the number of stages of the network.At the same time,the network module decomposed by CP is used to limit our coefficient matrix,so as to obtain better results.The experimental results show that the fusion method is effective on the three data sets. |
