Research On Hyperspectral Image Denoising Method Based On Total Variation And Low-Rank Tensor Decomposition

Due to various limitations of hyperspectral imaging process,hyperspectral images will degrade due to different noise sources.The existence of complex noise greatly reduces the quality of hyperspectral images,resulting in the failure of subsequent applications such as classification,unmixing,fusion,target detection,etc.Therefore,estimation of noiseless hyperspectral image is thus a critical element in hyperspectral remote sensing applications.The denoising method of hyperspectral images based on low-rank tensor theory and regularization method has become an significant method to remove the mixed noise of hyperspectral images.Designing an ideal denoising regularization of hyperspectral image with prior knowledge and to deeply explore and accurately describe the space and spectral knowledge in hyperspectral data is the core problem of hyperspectral image restoration.Based on theregularization theory and the intrinsic structural information of hyperspectral images,the main contributions of this thesis are as follows on the basis of the state-of-the-art techniques:（1）Aiming at the problem that the denoising model of hyperspectral low-rank tensor decomposition is easy to be excessively smooth due to the use of L₁ norm spatial-spectral total variational regularization in hyperspectral image mixed noise removal,a hyperspectral low-rank tensor mixed noise removal method based on the spatial-spectral hyper-laplacian total variational regularization is proposed.The hyper-laplacian distribution can approximate fit the heavy-tailed distribution of hyperspectral image gradients,a kind of spatial–spectral hyper-laplacian total variational regularization is designed.Compared with the L₁ norm,the non-convex L_p（0<p<1）norm has better sparsity,and the non-convex L_p（0<p<1）norm is used to describe the regularization.A hyperspectral low-rank tensor denoising model with spatial–spectral hyper-laplacian total variational regularization is established.The proposed denoising model is solved based on the Alternating Direction Method of Multipliers.Simulation and real experiment data of hyperspectral data show the superiority of the proposed method.（2）The imaging scenes of hyperspectral images are the same,and the local piecewise smoothness regions of different bands are also the same,a hyperspectral low-rank tensor mixed noise removal method based on the sparse total variational regularization of the joint group is proposed.Hyperspectral difference images have obvious row-group sparse structural priors when expanded along the spectral dimension,and a joint group sparse total variational regularization is designed.The designed associative group sparse total variational regularization is characterized by the L_2,p（0<p<1）norm.A hyperspectral low-rank tensor denoising model with joint group sparse total variational regularization is established.The proposed denoising model is solved based on the Alternating Direction Method of Multipliers.Simulation and real experiment data of hyperspectral data show the superiority of the proposed method.