Hyperspectral Image Denoising Via Hybrid Spatio-Spectral Total Variation Regularized Low-Rank Tensor Decomposition

Hyperspectral imaging technology is capable of providing rich spatial and spectral information at the same time,and has attracted a lot of attention in recent years.Due to the large amount of information it carries and its high spectral resolution,it can significantly improve the accuracy of object detection.Therefore,It is widely used in a variety of practical survey scenarios such as artwork identification,crop health,coastline mapping,mineral exploration,etc.,with a large scope for application development.However,hyperspectral image（HSI）is susceptible to noise（random noise）,blur（Gaussian and uniform blur）and downsampling（spectral and spatial downsampling）due to the limitations of hardware facilities and external environment.This causes significant inconvenience to subsequent image processing applications.Therefore,it is particularly important to obtain high quality HSI.In this paper,two HSI denoising models are proposed based on the inherent structural characteristics of HSI,both of which are able to remove complex noise well while retaining more detailed information.They perform well in experiments on simulated and real datasets.The specific work is as follows:（1）Based on the analysis of the sparsity of HSI structure,a hyperspectral image denoising model based on group sparsity regularized hybrid spatio-spectral total variation（GHSSTV）and low-rank tensor decomposition is proposed.First,the GHSSTV regularization method is proposed in this paper to ensure group sparsity not only on the first-order gradient domain but also on the second-order gradient domain along the spatial-spectral dimension.Secondly,Tucker decomposition is used in the model to study the global correlation of HSI,and low rank constraint is applied to it.Then,in order to efficiently remove complex noise,sparse noise in the HSI is detected using the L₁norm.Strong Gaussian noise is modelled using the Frobenius norm.（2）An HSI denoising model based on double domain low-rank constraint and hybrid space-spectral total variation（LRHSSTV）is proposed.To be specific,first of all,the low-rank property of the HSI gradient domain is proved using the transform domain low-rank theory,while the low-rank prior of the mixed gradient domain is ensured by the nuclear norm minimization.Secondly,on the basis of the first-rank gradient,the higher-rank gradient is introduced and the gradient tensors in the four directions of the spatial domain and the space-spectral domain are established,which can fully exploit the intrinsic correlation between spectral and spatial dimensions of higher order differences.Then,based on the prior structure characteristics of HSI,a weighted L_2,1norm is used to guarantee the gradient group sparsity priors of the mixed gradient tensor.Finally,the low-rank prior of the HSI original domain is verified by singular value decomposition,and the classical Tucker decomposition is used to impose low-rank constraints on the HSI original domain.（3）The alternating direction multiplier method（ADMM）algorithm is used to solve the above models step by step.The optimal solution of the model is obtained by constantly optimizing the subproblems in the model.The proposed model and the current most advanced model are tested on the simulation data sets and the real data sets,and the experimental results are compared to prove the superior performance of GHSSTV method and LRHSSTV method in the field of HSI denoising from both quantitative and qualitative analysis.Finally,the feasibility of the proposed model in practical application is further proved by classification experiment.