Research On Hyperspectral Image Fusion Method Based On Subspace Tensor Decompositio

Limited by the imaging mechanism of the sensor,the hyperspectral image has rich spectral information,but it needs to pay the price of corresponding spatial resolution.In practical applications,accurate structural information and rich spectral feature information are equally important,and single-sensor imaging is difficult to meet this requirement.At this stage,the best way to solve this problem is to use the imaging characteristics of different sensors to obtain complementary image information through fusion.For hyperspectral images,multispectral images with more spatial details can be selected as supplements.Therefore,this paper introduces the hyperspectral image fusion framework based on spatial and spectral degradation mechanisms,introduces the tensor decomposition method,deeply analyzes the properties of hyperspectral images,and makes a detailed summary of existing hyperspectral image fusion methods,proposes a new hyperspectral image fusion model.In addition,the effectiveness and high performance of the proposed model are verified through sufficient experiments on multiple datasets.The main research content is as follows:(1)A hyperspectral image fusion method based on low-rank learning and non-local priors is proposed.A non-local self-similarity prior is introduced to transform the fusion problem into the reconstruction of non-local tensor blocks.Through the tensor low-rank representation model,the low-rank property of the coefficient tensor in the non-local mode and spatial mode is explored in the low-dimensional subspace,and the sub-tensor coefficient is constrained by the truncated singular value low-rank approximation method,and the orthogonal constraint learning is used for basis and coefficients of the subspace.Through experiments on three simulated datasets and one real dataset,it is fully verified that the method effectively improves the accuracy and performance of fusion.(2)A fusion method based on spectral smoothing prior and sparse representation of tensor rows is proposed.This method follows the structure of non-local tensor blocks used in previous methods.However,unlike the former method that uses orthogonal constraints to learn subtensor dictionaries,we learn more continuous and smoother atoms that are closer to the real spectrum through continuous difference regularization.We also deeply explore the structure of the sub-tensor coefficients and accurately characterize the row sparsity of the sub-tensor coefficients to achieve efficient fusion of hyperspectral and multispectral images.Through experiments on three simulated datasets and one real dataset,it is fully verified that this method makes deeper use of the spatial and spectral properties of the target hyperspectral image itself,thereby improving the accuracy and performance of fusion.