Hyperspectral And Multispectral Image Fusion Based On Nonlocal Low-rank Representation And Its Application

Hyperspectral imaging combines imaging technology with spectral technology,and the obtained hyperspectral image is a data cube with three dimensions,which contains rich spectral information.Hyperspectral image plays an important role in many fields,such as land classification,food safety,geological survey and military monitoring,to name a few.However,due to the limitation of imaging equipment and solar radiation illumination,the imaging system cannot obtain spectral images with high resolution and many spectral bands,which greatly limits the application of spectral images in various fields.Recently,to obtain hyperspectral images with high resolution,the fusion of high-resolution multi-spectral images and low-resolution hyperspectral images has been widely studied.Due to the advantages of tensor in representing multi-order and multi-dimensional data,the use of tensor tools to deal with the fusion of hyperspectral and multispectral im-ages has gradually attracted people’s attention.Considering that spectral images can be naturally represented as third-order tensors,this thesis studies the fusion of spectral im-ages from the perspective of tensors.Aiming at the problems that it is difficult to explore the known spectral image information and the physical meaning is poor interpretable in the fusion of hyperspectral and multispectral images,this thesis uses tensor representation to mine the prior information of known images,and combines nonlocal low-rank repre-sentation to realize the fusion of spectral images.The main work of this thesis can be summarized as follows:1.A low transformed tensor rank spectral image fusion model based on nonlocal self-similarity is proposed.By analyzing the nonlocal self-similarity of spectral images,the clustering algorithm is used to find the similar sub-data cubes in the spectral images,and the similar data are rearranged into the form of a third-order tensor.The nonlocal low-rank structure in the images is further explored by using the rank of the transformation tensor.Finally,the fusion of spectral images is achieved by combining the idea of segmentation optimization and the ADMM to solve the objective optimization function.Experimental results show that the proposed model can achieve better fusion effect than the contrast method.2.A spectral image fusion model based on space-spectrum and low transformed tensor rank is proposed.On the basis of nonlocal self-similarity,by analyzing the spectral correlation between spectral bands,the constraint of low transformed tensor rank is simul-taneously applied to the reconstructed tensor and the rotation tensor of the reconstructed tensor,and the transform matrices involved in the transformed tensor rank are all constructed from known spectral images.Finally,the fusion of spectral images is achieved by solving the objective optimization function with the idea of segmentation optimization and the ADMM.The results of spectral image fusion experiment and ground object clas-sification application experiment show that the proposed model can obtain better fusion performance and classification effect when the number of spectral bands is higher.