CN107316309B - Hyperspectral image saliency target detection method based on matrix decomposition - Google Patents

Abstract

The invention provides a hyperspectral saliency target detection method based on matrix decomposition. The spectral gradient of the original hyperspectral image is calculated in the spectral dimension, the spectral gradient characteristic of the image is extracted, the adverse effect caused by illumination is eliminated, and meanwhile, an image characteristic matrix is constructed; then, matrix low-rank sparse decomposition is carried out on the image feature matrix to obtain a low-rank matrix corresponding to a background part and a sparse matrix corresponding to the salient object, so that the problem of nonuniformity of blocks in the salient object is avoided, and the operation amount can be reduced while the salient object is detected.

Description

Hyperspectral image saliency target detection method based on matrix decomposition

Technical Field

The invention belongs to the technical field of image processing, relates to a method for detecting a saliency target of a hyperspectral image, and particularly relates to a method for detecting the saliency target of the hyperspectral image based on matrix decomposition.

Background

The hyperspectral image is image data obtained by recording spectral information of various ground objects observed in a field of view by using an imaging spectrometer. With the gradual maturity of the hyperspectral imaging technology, the indexes of the imaging equipment such as spectral resolution, spatial resolution and the like are greatly improved. The problems of object detection, identification, tracking and the like which are mainly developed on conventional images originally are gradually extended to hyperspectral data. At present, relevant research on the hyperspectral image saliency target detection problem is in a development stage. The existing hyperspectral image saliency target detection method mainly adopts an Itti model, replaces color features with spectral features of a hyperspectral image, and enables the model to be suitable for the hyperspectral image. For example, the document "s.l. moan, a.mansouri, et al, science for Spectral Image Analysis [ J ]. IEEE Journal of selected Topics in Applied Earth requirements and Remote Sensing,2013.6(6): p.2472-2479" is to use Spectral information by projecting a spectrum into the CIELAB color space and performing Principal Component Analysis (PCA) using an Image. At present, in the existing method, pixels are used as a basic unit for significance estimation, and differences between different pixel spectrums are evaluated through means of principal component analysis, Euclidean distance, spectrum Angle (Spectral Angle) and the like, so that the significance of each pixel is measured. The main problem of the method for reflecting the significance of the whole image by the pixel significance lies in the phenomenon of nonuniformity of the significant image, wherein the edge response of an object is large and the internal response of the object is low, in the detection result. In addition, the existing methods all depend on a single model, and cannot eliminate the influence of brightness change in a hyperspectral image on spectral data and huge calculation amount caused by data scale. Therefore, it is urgently needed to break through the inherent thought in the existing hyperspectral image detection method and provide a new hyperspectral image saliency target detection method to solve the existing problems.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides a hyperspectral significance target detection method based on matrix decomposition. The idea of matrix decomposition is different from an Itti model and a region comparison method, the idea of matrix decomposition is to construct a sparse matrix and a low-rank matrix by utilizing similarity on a hyperspectral image space and continuity on a spectrum, and a significance target is just the sparse matrix in matrix decomposition. When a characteristic matrix to be decomposed is constructed, the influence of uneven brightness on data is eliminated by using the spectral gradient characteristic, and the calculation amount is reduced by utilizing a bottom layer constructed super-pixel structure. The traditional method utilizing the regional comparison idea causes the problem of nonuniformity in the salient object, and the method of the invention is based on matrix decomposition, and the uniformity in the salient object is realized by jumping out of the framework of the traditional method.

A hyperspectral image saliency target detection method based on matrix decomposition is characterized by comprising the following steps:

step 1: spectral gradient feature generation: for each pixel i of the hyperspectral image, according to

Calculating to obtain the spectral gradient characteristics, wherein the spectral gradient characteristics of all pixels form a spectral gradient characteristic data block X ═ g₁,g₂,…,g_n}^T(ii) a Wherein the content of the first and second substances,

the jth component characterizing the spectral gradient of pixel i,

representing the original spectral vector y corresponding to pixel i_iThe j component, i is 1,2, …, n, j is 2, …, p, n is the total number of pixels of the hyperspectral image, p is the number of wave bands of the hyperspectral image, and Δ λ is the wavelength difference of adjacent wave bands;

step 2: constructing an image feature matrix: representing the spectral gradient characteristic data block X obtained in the step 1 into a two-dimensional image characteristic matrix F with the size of m multiplied by n, wherein each column is a spectral gradient characteristic corresponding to one pixel, m is the dimension of the spectral gradient characteristic, and m is p-1;

and step 3: low-rank sparse matrix decomposition: performing low-rank sparse matrix decomposition on the image characteristic matrix F according to the following formula, and solving the image characteristic matrix F by using an alternative iteration multiplier method to obtain a sparse matrix S:

wherein L is a low rank matrix representing the background portion; s is a sparse matrix representing a salient object; λ is a weight coefficient; i | · | purple wind_*Is the kernel norm, | ·| luminance₁Is a 1 norm;

and 4, step 4: calculating a saliency map: and (4) carrying out normalization processing on the sparsity matrix S obtained in the step (3) to obtain a saliency map, namely a saliency target detection result.

The invention has the beneficial effects that: the spectral gradient is calculated on the spectral dimension of the original hyperspectral image, the spectral gradient characteristic of the image is extracted, the adverse effect caused by illumination is eliminated, and meanwhile, an image characteristic matrix can be constructed; the low-rank sparse decomposition is carried out on the image feature matrix to obtain the low-rank background part and sparse significance matrix, so that the problem of nonuniform blocks in a significance object is solved.

Detailed Description

The present invention is further illustrated by the following examples, which include, but are not limited to, the following examples.

The hyperspectral remote sensing image is of a cubic structure, the spatial dimension reflects the reflectivity of pixels corresponding to different positions on the ground on a certain sunlight wave band, and the spectral dimension reflects the relation between incident light and reflected light of the pixels at the certain positions on different wave bands. A hyperspectral image can be represented as a p x n data set Y_n＝{y₁,y₂,…,y_nIn which y_iAnd f, the original spectral vector corresponding to the pixel i is 1,2, …, and n is the total number of pixels of the hyperspectral image.

1. Spectral gradient feature generation

The spectral gradient refers to the ratio of the difference of every two adjacent components along the original spectral vector to the difference of the corresponding wavelengths. And the vector consisting of a series of spectral gradients is called a spectral gradient signature. The spectral gradient image is obtained by calculating the spectral gradient characteristics for each pixel, so that the extracted spectral gradient characteristics can maintain the spatial relationship of the original image.

Wherein:

the jth component characterizing the spectral gradient of pixel i,

representing the original spectral vector y_iI is 1,2, …, n, j is 2, …, p is the number of wavelength bands of the hyperspectral image, and Δ λ is the wavelength difference of adjacent wavelength bands.

For each in the hyperspectral imageOne pixel calculates its corresponding spectral gradient feature according to equation (2), and all the spectral gradient features form a spectral gradient data block X ═ g₁,g₂,…,g_n}^T. The spectral gradient feature can reduce the brightness difference caused by uneven illumination to a certain extent, so that the influence of the difference on the subsequent steps can be reduced.

2. Constructing an image feature matrix

And after the steps are completed, converting the obtained spectral gradient characteristic data block X into a two-dimensional image characteristic matrix F. Each column of the two-dimensional image feature matrix F is a spectral gradient feature corresponding to one pixel, the number of columns is the number n of pixels, and the number of rows is the dimension p-1 of the spectral gradient feature.

3. Low rank sparse matrix decomposition

Since spatial similarities in the background, especially pixels within the same salient object, all have local spatial similarities, salient objects can be detected using low-rank sparse matrix decomposition. Furthermore, there are a large number of similar spatial structures in natural scenes, and this spatial redundancy means that the background is a low rank floor. To distinguish background from salient objects, the following matrix decomposition model was used:

wherein: l is a low-rank matrix, S is a sparse matrix, is a weight coefficient and can take any rational number, and lambda is 3, | | · | | survival_*Is the kernel norm, | ·| luminance₁Is a 1 norm.

And decomposing the image feature matrix F into a low-rank background matrix L and a sparse detection matrix S, limiting the low-rank property of L by a nuclear norm, and limiting the sparsity of S by a 1 norm. Since the salient objects only occupy a small number of pixels compared with the whole image, the solution of the salient object can find the sparse matrix S by solving equation (3), and the limitation caused by the Itti model is well avoided.

For the objective equation (3), since it involves two variables S, L, which are separable objectives, it can be solved by using Alternating Direction Multiplier Method (ADMM), which is as follows:

first, an auxiliary variable H is introduced and equation (3) is written as:

then, introducing a lagrange multiplier distorts the function as:

wherein: l is_γ,η(L, S, H, P, Q) is the Lagrange equation, P, Q is the Lagrange multiplier, γ, η are penalty coefficients, | | · | | survival_FIs the F norm, const is a constant.

Next, the iterative update is performed according to the following steps:

(1) fixing other parameters, the matrix L is updated as follows:

wherein, U Σ V^TIs H^k+Q^kThe result of the singular value decomposition of/η,

representing the soft threshold of the calculation matrix sigma.

(2) Fixing other parameters, updating the matrix S as follows:

wherein the content of the first and second substances,

representing the operation of computing the matrix soft thresholds.

(3) Fixing other parameters, the matrix H is updated as follows:

where I is the identity matrix.

Repeating the steps (1) to (3) to carry out iterative updating until the iteration number k exceeds the specified maximum number, or

And (e is more than or equal to 0), finishing the iterative updating to obtain the sparse matrix S.

4. Saliency map computation

And (4) carrying out normalization operation on the sparsity matrix S obtained in the above step to obtain a saliency map, namely a saliency target detection result.

Experiments prove that the precision ratio can be improved by 10% when the recall ratio is 0.7; when the precision ratio is 0.7, the recall ratio can be improved by 20 percent.

Claims (1)

1. A hyperspectral image saliency target detection method based on matrix decomposition is characterized by comprising the following steps:

step 1: spectral gradient feature generation: for each pixel i of the hyperspectral image, according to

Calculating to obtain the spectral gradient characteristics, wherein the spectral gradient characteristics of all pixels form a spectral gradient characteristic data block X ═ g₁,g₂,…,g_n}^T(ii) a Wherein the content of the first and second substances,

the jth component characterizing the spectral gradient of pixel i,

representing the original spectral vector y corresponding to pixel i_iI-1, 2, …, n, j-2…, wherein p and n are the total number of pixels of the hyperspectral image, p is the number of wave bands of the hyperspectral image, and delta lambda is the wavelength difference value of adjacent wave bands;

step 2: constructing an image feature matrix: representing the spectral gradient characteristic data block X obtained in the step 1 into a two-dimensional image characteristic matrix F with the size of m multiplied by n, wherein each column is a spectral gradient characteristic corresponding to one pixel, m is the dimension of the spectral gradient characteristic, and m is p-1;

and step 3: low-rank sparse matrix decomposition: performing low-rank sparse matrix decomposition on the image characteristic matrix F according to the following formula, and solving the image characteristic matrix F by using an alternative iteration multiplier method to obtain a sparse matrix S:

wherein L is a low rank matrix representing the background portion; s is a sparse matrix representing a salient object; λ is a weight coefficient; i | · | purple wind_*Is the kernel norm, | ·| luminance₁Is a 1 norm;

and 4, step 4: calculating a saliency map: and (4) carrying out normalization processing on the sparsity matrix S obtained in the step (3) to obtain a saliency map, namely a saliency target detection result.