Research On Hyperspectral Remote Image Reconstruction Based On Low-Rank Constraint

Hyperspectral image(HSI)contains rich spatial information and spectral information of the target terrain,and its spectral resolution is up to nanometer level.It is widely used in geological exploration,agroforestry,ocean remote sensing,military surveillance,and many other fields.However,in the process of imaging and transmission,due to various factors such as atmospheric influence and sensor failure,the acquired hyperspectral image is often degraded by the influence of noise.The image quality is degraded,which may bring great challenges to the post application such as image classification and target recognition.So studying how to remove noise from hyperspectral images and recovering clean hyperspectral images is a direction worthy of further study.In the initial stage of research on hyperspectral image restoration methods,scholars usually regard hyperspectral images as a collection of two-dimensional images or one-dimensional signals,and denoise them by using mature image or signal recovery methods one by one.Those methods ignore the correlation between adjacent bands or adjacent pixels in the image.Thus,the image restoration effect is poor.Later,the researchers regarded the hyperspectral image as a whole,and proposed a hybrid space-spectral derivative domain wavelet contraction model and a multidimensional Wiener filter to recover,but the effect is still not satisfactory.In recent years,researchers have explored the characteristics of hyperspectral images from various angles.By utilizing regularization constraints,many constrained hyperspectral image restoration models have been proposed.Such as Low-Rank Matrix Recovery(LRMR)models,Spatial Adaptive Total Variation(SATV)model and Total-variation-regularized Low-rank Matrix Factorization(LRTV)model.This paper focuses on hyperspectral image restoration methods based on sparse and low rank constraints.The goal is to recover clean hyperspectral images from degraded hyperspectral images to meet the needs of subsequent applications.The recovery problem of hyperspectral image can be expressed as obtain the image degradation model based on the image degradation,and then solve the inverse problem of degradation,establishing an effective recovery model based on image characteristics,which is the core of this paper.Firstly,the reason for the degradation phenomenon of hyperspectral image is analyzed,and a reasonable recovery model and optimization equation are established.Then,according to the characteristics of hyperspectral image,the low-rank and sparse regular term is used to limit the solution space of the optimization equation.When solving the recovery problem,the solution process is more complicated because there are more than one variable.In order to reduce the computational complexity while ensuring the optimal results of the algorithm,this paper mainly uses the Augmented Langrange Method(ALM)and Maximum-Minimize(MM)to solve the target optimization problem effectively.The purpose of hyperspectral image restoration is to improve image quality,facilitate the extraction of useful information,and improve the accuracy of subsequent applications.This paper mainly studies hyperspectral image restoration methods based on low rank and sparse constraints.Firstly,the hyperspectral image restoration model based on convex constraints is established by using the low rank and smoothness of hyperspectral images,including hyperspectral image restoration model based on low rank constraint and hyperspectral image restoration model based on low rank total variation constraints.Secondly,considering the existence of estimation bias based on convex constraint-based restoration methods,the hyperspectral image restoration method based on nonconvex low rank constraint is studied,including hyperspectral image restoration method based on ? norm low rank constraint and hyperspectral image restoration method based Schatten-p norm lowrank constraint.The recovery effect of these methods is improved compared with the convex constraint based recovery method.Finally,two hyperspectral image restoration methods based on nonconvex low rank sparse constraints are proposed.The hyperspectral image restoration method based on nonconvex low rank constraint can accurately approximate the rank of the real image,thereby effectively improving the quality of the restored image,but lacking the utilization of spatial information.In order to make full use of the low rank and smoothness of hyperspectral images,this paper innovatively proposes a nonconvex low rank sparse constrained hyperspectral image restoration method based on MM algorithm and a nonconvex low rank sparse constrained hyperspectral image restoration method based on ALM algorithm.The simulation results show that compared with the hyperspectral image restoration method based on nonconvex low rank constraint,the proposed method can not only effectively remove noise,but also retain more specific details.The image recovery quality is higher.The purpose of effectively recovering clean hyperspectral images from the degraded hyperspectral images is achieved.