Research On Remote Sensing Hyperspectral Image Denoising Based On Tensor Low Rank Sparse Restoration Theory

Remote sensing hyperspectral imaging is a kind of multi-band image data which combines two-dimensional spatial information and spectral information.It can acquire the continuous narrow band image data information,these images have the high resolution characteristic,may provide the rich ground object detail and the characteristic,therefore,it is widely used in geological exploration,vegetation ecological monitoring,atmospheric Environmental monitoring,agricultural remote sensing,marine survey and other applications.However,since the hyperspectral imaging is passive and the remote sensing platform is far from the ground,during the image acquisition process,can be affected by a variety of factors such as sensor system hardware,random failure of sensor components,non-uniform light intensity,and atmospheric turbulence,image quality problems such as low spatial resolution,pixel loss,stripe noise,mold and burn,uneven brightness distribution,shadow,cloud occlusion and so on often occur in the acquired image,for example,gauss noise,random noise and fringe noise are the types of noise,which make the application of hyperspectral image classification difficult.In this paper,the fundamental problem of removing the mixed noise from remote sensing hyperspectral images is studied,and the spatial and spectral information of the hyperspectral images and the low rank and sparse representation of the Hyperspectral imaging are combined with tensor decomposition,a series of de-noising methods for hyperspectral images are proposed.The main research work and achievements of this paper are as follows:?1?An adaptive total variational regularization?TSDM?model based on minimum maximum nonconvex penalty norm?MCP?constraint and one-way Tchebichef distance difference is proposed.Aiming at the problem that the Algorithm based on total variation regularization can not remove the mixed and irregular stripe noise effectively,and the image details are lost in the process of removing the stripe noise,in this paper,a new model,TMCP-SDM,is proposed by studying the inherent orientation and structure of stripe noise.Firstly,the image layer is smoothed in the vertical direction of the stripe noise and has a typical structure.Therefore,Tchebichef distance UTV is used to suppress the smooth subspace,which can effectively preserve the structure information of the image and reduce the step effect of the estimation image.Secondly,the model uses non-convex MCP-norm to represent the low rank of the stripe noise,because the non-convex approximation method is more accurate than the convex approximation methods such as l₁-norm and kernel norm instead of l₀-norm and rank function;it can reflect the low rank and sparsity of the stripe noise components.In the TMCP-SDM model,the minimum maximum nonconvex penalty regularization constraint of Tchebichef distance and the low rank approximation estimate of the image are constructed as prior information for recovering the low rank hyperspectral images and the stripe noise components,respectively,based on the ADMM optimization model,a method is proposed to calculate the restoration image and the stripe noise components,which can effectively separate the stripe noise from the noisy images.?2?A new hybrid noise cancellation method?Nonconvex Low Rank Tensor approximation and Phase Consistency for mixed denoising,NLRTAPC?based on nonconvex low rank tensor approximation and phase consistency is proposed.Aiming at the shortcomings of TV regularization and low-rank tensor de-noising methods based on band division,which can not effectively utilize local neighborhood information and can introduce ladder effect,it is easy to cause"artificial artifact"phenomenon,especially in the curved edge,and gradient information can not accurately describe the true structure of the image.Therefore,a new nonconvex smooth rank approximation model is proposed to deal with the mixed noise of hyperspectral images.The normalized Pierre-Simon Laplace function nonconvex low rank matrix is used to approximate the kernel norm,a new phase congruency l_p-norm model is proposed to constrain the spatial structure of hyperspectral images by fusing l_p-norm constraints of phase congruency constraints,Because the non-convex low-rank approximation performs better in low-rank and sparse structure,the"artificial artifact"phenomenon in hyperspectral image denoising can be well solved.?3?A low rank hybrid tensor approximate decomposition model HLRTD and a weighted space and spectral TV SSTV total variation regularized tensor low rank hybrid decomposition model,HLRTD-SSTV,are proposed.In order to solve the problem that the existing methods based on low-rank tensor decomposition for hyperspectral image denoising can not extract the fine-grained sparsity within the tensor,this paper proposes a low-rank tensor approximation,by minimizing both CP-rank and Tucker-rank simultaneously,the low rank structure and sparsity of tensor can be accurately described by both CP rank and Tucker rank,which is more favorable for image restoration.In this paper,a low rank hybrid tensor low rank decomposition model?Hlrtd?is proposed to describe the low rank structure and sparse information of HSI.In order to make full use of the high-dimensional structure information in hyperspectral images,the low-rank hybrid tensor approximate decomposition model?HLRTD?decomposes the CP model rank and the Tucker model rank to approximate the natural structure of hyperspectral images using the low-rank approximation,the weighted space and total variation regularization of spectral TV?SSTV?is introduced into the tensor low-rank mixed decomposition model to ensure the restored image has a good local smooth structure and additional spectral complementary information when the mixed noise is eliminated,to reduce spectral distortion.By using alternating direction multiplication ADMM,the optimization model of two tensor ranks is transformed into two sub-problems,each sub-problem has only one tensor rank.For the two decomposed subproblems,the tensor low rank approximation updating method and the matrix weighted kernel norm minimization method are used respectively.From the results of simulation experiments and real data experiments,it is found that the HLRTD-SSTV model can effectively deal with gauss noise,striping noise and mixed noise,etc.,all showed good index results.?4?Three models,TSDM,NLRTAPC and HLRTD-SSTV,were integrated and applied to real hyperspectral datasets,the three models were applied to the real hyperspectral digital image HYDICE Urban dataset by visual and quantitative evaluation methods,and the validity of the models was further verified.