Research On Subspace Low-rank Learning Methods For Hyperspectral Image Denoising

Hyperspectral remote sensing combines the spatial information of the ground objects with the extremely rich spectral information,which greatly expands human vision and enhances the ability of substances detection.In the growing applications of hyperspectral images,the accuracy of information extraction and processing largely depends on the quality of the collected data.However,during the process of signal sampling and digital imaging,hyperspectral images are inevitably degraded by various mixed noises.This degradation can greatly reduce the accuracy and efficiency of subsequent applications.Recently,low-rank representation-based methods have drawn a lot of attention in the community of image inverse problems.However,hyperspectral image data possess extremely high dimension and redundancy,hence it is difficult to fully utilize its spatial and spectral prior through directly employing the low-rank matrix recovery,which leads to insufficient denoising accuracy.Meanwhile,because hyperspectral images have many bands and a large volume of data,lowrank representation methods based on model optimization are often inefficient.To address the above challenges,in this thesis,we first took a closer look at the self-portrait of hyperspectral image and studied existing denoising methods in detail.On this basis,we designed new restoration models,proposed a series of high-performance methods for hyperspectral images denoising,and verified their effectiveness through extensive experimental evaluations.The main contributions of this thesis are summarized as follows:(1)We proposed a novel joint method of subspace low-rank learning and non-local 4-d transform filtering for hyperspectral image denoising.Unlike the previous low-rank restorations that generally impose the low-rank constraints on the original hyperspectral image data,the proposed method characterized the spatial-spectral low-rank nature of the hyperspectral image in a low-dimensional subspace and coupled the parameter-free BM4 D denoiser as a regularization to take advantage of the non-local self-similarity of the coefficient image.Extensive experiments verified that the proposed method not only improves the accuracy and efficiency of denoising but also improves practicality.(2)We proposed a tensor subspace low-rank learning method with non-local prior for hyperspectral image denoising.Unlike the previous subspace representation methods that generally unfold the image tensor into a two-dimensional matrix before implementation,the proposed method realized a fully tensorized subspace low-rank learning.Through the highorder linear representation,the proposed method learns the basis tensor and coefficient tensor in subspace.Extensive experiments verified that the proposed method not only maintains the high efficiency of the subspace-based method but also further improves the denoising accuracy by successfully maintaining the high dimensional structure correlations of image tensor.(3)We proposed a difference continuity-regularized tensor subspace low-rank learning method with successive low-rank tensor decomposition for hyperspectral image denoising.Unlike previous subspace low-rank learning methods that generally use an orthogonal constraint and a plug-and-play denoiser to learn the basis and coefficient of subspace respectively,the proposed method learns a more continuous and smoother spectrum basis through difference continuity-regularization while exploiting the low-rankness of the coefficient tensor along the spectral,nonlocal,and spatial modes through successive low-rank tensor decomposition.Extensive experiments verified that the proposed method makes more comprehensive utilization of the intrinsic spatial-spectral characteristics of the hyperspectral image,which further improves the mixed denoising accuracy.