Research On Hyperspectral Remote Sensing Image Restoration Based On Low-rank And Sparse Constraint

Hyperspectral remote sensing images can provide both spatial and spectral information for ground features,which not only reflect the size,shape and orientation of the ground features,but also provide information including material,physical structure,chemical composition,etc.They can greatly improve the accuracy and reliability of remote sensing monitoring and anal-ysis,making hyperspectral remote sensing imaging technology widely used in environmental monitoring,urban planning,agricultural and forestry detection,land survey and other fields.In the process of hyperspectral data acquisition,compression and transmission,due to the in-fluence of instrument itself,external environment,etc,the acquired hyperspectral image（HSI）is often degraded,which seriously affects the performance of subsequent object information extraction.Therefore,how to recover the original HSI from degraded image is highly related to the accurate extraction and interpretation of ground object information,and is of great sig-nificance to the subsequent HSI applications.This thesis mainly studies HSI restoration tech-nology,including the inpainting,denoising and compressive sensing reconstruction,based on the degradation mechanism and internal structure characteristics of hyperspectral image.The proposed methods combine machine learning and optimization frontier related theories,and the low-rank and sparse constraint models and numerical optimization algorithms are constructed.The main work of this thesis is as follows:（1）Aiming at the problem of loss of high-dimensional features caused by compressive sensing reconstruction methods transforming high-dimensional data into low-dimensional data,an HSI compressive sensing reconstruction method with tensor low-rank and local smooth-ness is proposed in this thesis.The proposed method treats the HSI as a whole,and the HSI spatial-spectral global correlation are characterized by tensor low-rank decomposition,and the local correlation between adjacent space and adjacent spectra are characterized by anisotrop-ic space-spectral total variation regularization to better maintain the edge information of the image.By combining the features of hyperspectral data spatial and spectral correlation with local piecewise smoothing,an HSI compressive sensing reconstruction method is introduced,which preserves the high-dimensional structure characteristics of the image and can recover the original HSI more accurately.In experiments,the proposed method is compared with sever-al main hyperspectral compressed sensing reconstruction methods.The experimental results verify that the proposed method can effectively improve the accuracy of compressive sensing reconstruction.（2）To solve the problem of missing hyperspectral pixels,large area and continuous bad pixels,this thesis introduces an HSI inpainting method via tensor low-rank regularization and nonlocal self-similarity.In the proposed method,the degraded HSI is initially repaired by the constructed tensor low-rank completion model,and the initial inpainting result is obtained.Then,by exploiting the HSI nonlocal spatial-spectral similarity property,similar patches in the initial inpainting result are aggregated into a class,and each class is constrained by the con-structed tensor low-rank completion model,and the final inpainting result is achieved by inte-grating the inpainting result of each class.On the one hand,since the tensor trace norm shrinks each singular value equally,and the large singular value corresponds to the main information in the image,and thus we should shrink the larger singular value less,and treat each singular val-ue adaptively.By combining the log det（X）with the tensor trace norm,the constructed tensor rank surrogate function is able to treat each singular value adaptively,which can improve the flexibility of the tensor trace norm and tightly approximate the tensor rank.On the other hand,since similar patches have similar structure information,by grouping similar patches into one class,the resulting 3D data has better low-dimensional character,which can further improve the texture and detail restoration accuracy as well as inpainting performance.In experiments,the proposed method is compared with multiple mainstream tensor completion methods.In terms of signal-to-noise ratio and visual effect,the proposed method can accurately restore the HSI missing elements,and greatly improve the image restoration accuracy.（3）Aiming at the problem of the HSI degradation caused by complex noise（including Gaussian noise,impulse noise,stripe,etc.）,this thesis introduces a nonconvex low-rank and s-parsity regularized HSI mixed noise removal method.The proposed method aims to decompose the degraded HSI into clean HSI,sparse noise and Gaussian noise components.A nonconvex regularizer,i.e.,normalizedε-penalty is introduced to enhance the sparsity of clean HSI and the sparse noise.The proposed normalizedε-penalty can adaptively shrink each entry,and thus enhance the sparsity in both the intrinsic low-rank structure of clean HSI and the sparse cor-ruptions.Besides,we develop an effective algorithm based on the majorization minimization（MM）to solve the resulting nonconvex optimization problem.Compared with other state-of-the-art HSI restoration approaches,experimental results show that the proposed method can effectively eliminate Gaussian noise,impulse noise,dead lines,and strips,and achieve high fidelity reconstruction of spatial-spectral structure information.（4）In order to solve the problem that the HSI is contaminated by complex noise and the nonconvex optimization problem is generally more challenge to solve compared with the con-vex optimization problem,the adaptive rank and structure sparsity corrections for HSI mixed noise removal is introduced in this thesis.The proposed method introduces two convex regular-izers,namely the rank correction and the structure sparsity correction,to respectively impose low-rank and structure sparsity constraints on the decomposed clean HSI and sparse noise.The rank correction can adaptively offset the penalty of the nuclear norm for large singular values,so it can better exploit the low-dimensional characteristic of clean HSI.The structure sparsity cor-rection is capable of adaptively offsetting the penalization of large entries from the L_2,1-norm,and thus enhance the structure sparsity of sparse noise.Therefore,the proposed method has the tighter approximations of the matrix rank and the L_2,0-norm.Besides,since the constructed regularizers are convex,the resulting optimization problem is a convex optimization problem,which enables itself to be easily solved and obtain a closed solution.Compared with other clas-sic high-dimensional image denoising methods and HSI denoising methods based on low-rank approximation,experimental results show that the proposed method can effectively eliminate complex noises in degraded HSI and achieve high fidelity reconstruction of HSI spatial-spectral structure information.