Hyperspectral Image Denoising Algorithm Based On Low-Rank Representation

With the maturity of hyperspectral remote sensing imaging technology,hyperspectral image has attracted widespread attention in remote sensing applications due to the large amount of available spectral information.In particular,it is higher than the classification and recognition capabilities of traditional remote sensing images,and is of great significance in the classification of features,target detection,urban planning,etc.However,the collection and transmission of hyperspectral image data are inevitably affected by many factors.This generated noise seriously affects the image quality,which also brings difficulties to the refined applications of hyperspectral image.Therefore,as an important preprocessing step,the research of hyperspectral image denoising is very important,and its denoising performance will affect the accuracy of subsequent processing.Based on this,this thesis focuses on the low-rank property of hyperspectral image and studies denoising algorithms.Specifically,it includes the following two parts:First,we propose a hyperspectral image denoising algorithm based on Nonconvex constrained Low-Rank Sparse Decomposition.Based on the underlying low-rank property of hyperspectral image,joint denoising from spatial and spectral dimensions through sparse constraint and non-convex penalty function.The model is solved by the augmented Lagrange multiplier method.Experimentally,the effectiveness of Nonconvex constrained Low-Rank Sparse Decomposition algorithm is verified by simulated noise experiments and real data experiments.Second,we propose a hyperspectral image denoising algorithm based on the Total Variation constrained Low-Rank Sparse Decomposition.By performing low-rank constraint to explore the intrinsic spatial-spectral correlation of HSI,and simultaneously recovering the clean low-rank HSI from the observed contaminated HSI data.In order to better separate the useful signal from noise,TV regularization and sparse regularization are combined to process the noise component.The experimental results confirm that the Total Variation constrained Low-Rank Sparse Decomposition algorithm clearly improves the denoising results,and is superior to other methods in terms of visual and quantitative evaluation.