| In recent decades,optical remote sensing images has moved from the stage of black-and-white panchromatic images into the new stage of hyperspectral images.Compared with the traditional multi spectral images,hyperspectral images have many characteristics such as more bands,higher spectral resolution and low spatial resolution.Hyperspectral images,as three-dimensional image cube with the combination of image and spectral,have not only two-dimensional description information of object distribution,but also the one-dimensional spectral information feature attributes.For military reconnaissance,maritime search and rescue,geological exploration rare minerals,medical diagnosis of cancer cells and so on,the hyperspectral images are one of the important data.Hyperspectral image anomaly detection is one of the most important applications.The advantage of hyperspectral image anomaly detection is that without the help of spectral library and inversion algorithm,it can accomplish the extraction of outliers.Based on the characteristics of low rank and sparse,and combined with low dimensional space structure and space distance constraint,this article will propose two algorithms:First of all,this article proposes an anomaly detection algorithm based on latent low rank representation model and Laplacian matrix constraint.There are the same type of background pixels in hyperspectral images,and their spectral curves are similar.Therefore,the weight coefficients corresponding to these similar background pixels are large in the representation.While the other types of background pixels contribute only a few weights,and the representation coefficient will be in the low rank space.In the low rank representation,the construction of the background pixel dictionary will affect the reconstruction of background pixel in hyperspectral images.The background pixel dictionary is constructed by using the pixels in the hyperspectral image to extract the main features of the background pixels,and the anomaly pixels are separated from the hyperspectral image.Because of the high dimensional geometric structure of hyperspectral image data,the Laplasse matrix is introduced to restrict the low dimensional local structure in the space,and get more accurate representation coefficient.Then this paper proposes another fast soft threshold iteration hyperspectral anomaly detection algorithm based on sparse representation and space distance constraints.The spectral features of background pixels in the hyperspectral image are reconstructed by linear combination of atoms in the background dictionary.In the background dictionary only a small part of the atoms participate in it.So the use of norm constrains the sparse characteristic of representation coefficients.At the same time the distance weight between the test pixel and the atoms in the background dictionary as a constraint improve the ability of representation.If the space distance of the test pixel space and the atom is further,the contribute of atom is smaller.The improvement of spectral and spatial dimensions of hyperspectral imagery increases the amount of data.In order to meet the requirements of real-time,the efficient and simple fast soft threshold iterative algorithm is used to solve the target equation,which can improve the detection speed. |
