Research On Hyperspectral Image Denoising Methods Based On Tensor Low-Rank And Sparse Priors

Hyperspectral remote sensing is the frontier field of remote sensing technology.It obtains hyperspectral image(HSI)data by combining imaging technology and spectral technology.The obtained hyperspectral image data has the characteristics of many bands,high spectral resolution and high redundancy,and can provide very rich features of ground objects,so it is widely used in many application fields such as environmental monitoring,agricultural remote sensing,geological exploration and military reconnaissance.However,because the imaging spectrometer is easily influenced by its own equipment and external environmental factors in the imaging process,the obtained hyperspectral images are usually interfered by various mixed noises(such as,Gaussian noise,salt and pepper noise,stripes and deadlines,etc.),which not only seriously affects the image quality,but also greatly affects the subsequent processing and applications.As an important and effective scheme for the quality improvement of HSI,the HSI denoising problem aims to recover a clean HSI from the HSI with mixed noises.Therefore,for the problem of HSI denoising,this paper studies the removal of mixed noise and stripe noise in HSI.Based on the tensor representation modeling of HSI,this paper focuses on the modeling of the potential tensor low-rank characteristics of HSI,the sparse characteristics of sparse noise and the low-rank prior characteristics of stripe noise,and proposes two denoising methods for HSI based on tensor low-rank and sparsity priors.The main research work of this paper is listed as follows:(1)A hyperspectral image denoising method based on spatial and spectral gradient tensorsbased low-rank priors is proposed.Firstly,the low-rank characteristics of HSI in spectral gradient domain are studied,and the spectral gradient low-rank prior term based on weighted nuclear norm is proposed.Secondly,the spatial-mode low-rank characteristics of spatial gradient tensors of HSI in spatial gradient domain are studied,and the spatial gradient tensor low-rank prior term based on tensor nuclear norm is proposed.Finally,combined with the tensor sparse prior of sparse noise,a unified hyperspectral image denoising model based on the spatial and spectral gradient tensors-based low-rank priors is proposed,and an optimization algorithm for solving the model is designed.(2)A hyperspectral image destriping and denoising method based on non-local and global lowrank priors in tensor subspace is proposed.Firstly,the spectral-dimensional tensor low-rank property of HSI is modeled by using the tensor nuclear norm.Secondly,the non-local self-similarity of the low-rank representation coefficient tensor in the tensor subspace of HSI is studied,and a non-local low-rank prior term is imposed.Then,for the modeling of special properties of stripe noise,the lowrank characteristics of stripe noise are studied,and then a global low-rank prior term for stripe noise based on nuclear norm is proposed.Finally,combined with the tensor sparse prior term of sparse noise,a unified hyperspectral image destriping and denoising model based on tensor subspace-based non-local and global low-rank priors is proposed,and an optimization algorithm for solving the model is designed.Furthermore,this paper conducts a lot of experiments on the simulated and real HSI datasets,and compare the proposed methods with various mainstream HSI denoising methods.The experimental results show that the proposed methods are superior to many mainstream methods in terms of visual and quantitative comparisons,and show better visual denoising effect and relatively better results of MPSNR,MSSIM and SAM.