Global anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method

Abstract

The application discloses a hyperspectral remote sensing image constraint collaborative representation anomaly detection method based on global anomaly weighting. A local dictionary is constructed by using a sliding double window, a graph regular term constraint is added on the basis of collaborative representation in order to maintain the local geometric structure in the original data space, a graph constraint collaborative representation model is established, and a global anomaly saliency weight is constructed by using a spectral residual method in order to fully utilize the global features of the hyperspectral remote sensing image. Finally, the established graph constraint collaborative representation model and the constructed global anomaly saliency weight are nonlinearly fused. The application aims at solving the problems of insufficient utilization of the best representation of the to-be-tested image element in the reconstruction model and features, and effectively improves the precision of the hyperspectral anomaly detection by using the graph constraint collaborative representation model and the global anomaly saliency weight.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral remote sensing image processing technology, specifically relating to a globally weighted hyperspectral remote sensing image constraint collaborative representation anomaly detection method. Background Technology

[0002] Hyperspectral remote sensing technology combines imaging and spectral techniques, enabling the simultaneous acquisition of spatial and spectral information of ground features, resulting in a unified hyperspectral image. The main difference between hyperspectral and traditional multispectral images lies in narrowband imaging. A hyperspectral image can be viewed as a three-dimensional data cube with two spatial dimensions and one spectral dimension. It records not only the spatial information of ground features but also their continuous spectral information. Different ground features have different materials, varying reflectance and absorption characteristics, resulting in different spectral curves. Therefore, hyperspectral images offer excellent opportunities for detailed identification of ground features. Because hyperspectral images contain a wealth of information and possess multiple bands and high spectral resolution, they can solve many problems that optical or multispectral images cannot, especially in identifying ground features based on the spectral characteristics of different materials. Hyperspectral images have unparalleled advantages in this area, but this also brings challenges such as the curse of dimensionality and data redundancy. Currently, hyperspectral imagery has significant application value and broad development prospects in fields such as precision agriculture, environmental monitoring, ground cover classification, and target detection.

[0003] Target detection refers to the detection and identification of ground objects of interest (GMOs) or small ground objects that differ from most ground object types in hyperspectral imagery. Based on whether prior information about the GMOs is known, hyperspectral target detection can be divided into supervised target detection with known prior information and unsupervised target detection with unknown prior information. Unsupervised target detection, also known as anomaly detection, is challenging because prior spectral information for many ground object types is difficult to obtain in real-world scenarios. Furthermore, atmospheric absorption, illumination, and sensor signal influences during atmospheric transmission lead to variability in the obtained spectra, resulting in phenomena such as different spectra for the same object and different objects with the same spectrum. This hinders target detection that requires prior information. In this context, anomaly detection, which requires no prior information, has significant research value. Because it does not require prior knowledge, hyperspectral anomaly detection technology is currently widely used in civilian search and rescue, national defense security, and medical examination.

[0004] In hyperspectral remote sensing anomaly detection, anomalous targets typically exhibit low probability of occurrence, small number, small area, and spectral features significantly different from surrounding pixels. Existing hyperspectral anomaly detection algorithms can be broadly categorized into three types: statistical model-based anomaly detection algorithms, nonlinear anomaly detection methods, and representation learning-based anomaly detection algorithms. The Reed-Xiaoli (RX) anomaly detection algorithm is considered the most classic algorithm in statistical models. This method assumes that background pixels follow a multivariate Gaussian distribution and identifies pixels deviating from the multivariate Gaussian distribution characteristics of the background features as anomalous targets. However, statistical models make a simplistic background assumption and fail to consider the diverse and unevenly distributed types of land features in actual hyperspectral imagery. Nonlinear anomaly detection algorithms map the background and anomalous data, which are originally difficult to distinguish linearly, to a higher-dimensional feature space, increasing the separability between data. The SVDD algorithm is a support vector method that uses a kernel approach. It avoids pre-assuming the distribution of the original background data and models the background using a minimum closed hypersphere, identifying spectral pixels exceeding the hypersphere model as anomalous. While nonlinear anomaly detection methods can improve the separability of data, mapping data to a higher-dimensional space introduces problems such as increased data volume and computational complexity. In recent years, representation-based methods without any statistical assumptions have attracted widespread attention. Representation-based anomaly detection algorithms assume that normal background pixels can be well linearly represented by a dictionary, while anomalous pixels cannot. Depending on different constraints, representation-based anomaly detection can be categorized into sparse representation, cooperative representation, and low-rank representation. Representative algorithms include BJSR, CRD, and LRASR. Representation-based anomaly detection algorithms have become a research hotspot due to their simplicity, strong generalization ability, and high efficiency.

[0005] Anomaly detection methods based on representation models effectively model the background, improving the accuracy of anomaly detection and offering advantages such as high computational efficiency and strong generalization ability. However, hyperspectral anomaly detection based on representation models still faces some challenges. Existing collaborative representation models suffer from rigid and singular constraints, and the dictionary atoms constructed using the sliding dual-window model are affected by unknown anomaly pixels, all of which impact the reconstruction performance. Furthermore, the dual-window model underutilizes the features of hyperspectral images, further hindering anomaly detection. Therefore, improving the reconstruction performance under representation models and fully leveraging the features of hyperspectral images to enhance anomaly detection are pressing issues that need to be addressed. Summary of the Invention

[0006] Purpose of the invention: This invention provides a globally weighted hyperspectral image-constrained collaborative representation method for anomaly detection, which can preserve the local geometric structure in the dictionary atomic space and improve the reconstruction effect of hyperspectral images under the representation model, making the anomalous parts more prominent, thereby further improving the detection accuracy of anomalous targets.

[0007] Technical Solution: To achieve the above-mentioned objectives, the global anomaly-weighted hyperspectral image-constrained collaborative representation anomaly detection method of the present invention includes the following steps:

[0008] Step 1: Convert the hyperspectral remote sensing image into a two-dimensional matrix X∈R b×n Where b is the number of bands and n is the number of pixels;

[0009] Step 2: For each pixel, construct a local dictionary set X using a sliding double window. s , where s is the number of atoms in the dictionary;

[0010] Step 3: Calculate the adjacency matrix W between dictionary atoms, and obtain the symmetric normalized Laplacian matrix L based on the adjacency matrix W. sym The obtained symmetric normalized Laplacian matrix is ​​used to construct a graph regularization term. The graph regularization term is used as one of the constraints to add a constraint term to the optimized collaborative representation model. The graph-constrained collaborative representation model is constructed, and the closed-form solution of the coefficient vector is obtained to obtain the initial anomaly detection results. The global anomaly significance weight of the hyperspectral image is calculated using global information and the spectral residual model.

[0011] The global anomaly significance weights of hyperspectral images are constructed using a spectral residual model. The specific steps are as follows:

[0012] (a) The optimal clustering framework (OCF) method is used to select bands in the hyperspectral remote sensing image. After obtaining the image I(x) generated by the three bands with the richest information, the grayscale image of the image is obtained.

[0013] (b) Perform a Fourier transform on the grayscale image, and then perform a logarithmic transform on the grayscale image A(f) after the Fourier transform to obtain the logarithmic amplitude spectrum L(f).

[0014] (c) Convolve the logarithmic spectrum L(f) using a local averaging filter to obtain the average logarithmic amplitude spectrum.

[0015] (d) Subtracting the mean logarithmic amplitude spectrum from the logarithmic amplitude spectrum yields the spectral residual R(f);

[0016] (e) Perform an inverse Fourier transform on the spectral residual R(f) and smooth it using a Gaussian filter G(x) to obtain the global anomaly significance weights.

[0017] Step 4: Nonlinearly fuse the initial anomaly detection results and the global anomaly significance weights;

[0018] Step 5: Output the final anomaly detection results.

[0019] Furthermore, the adjacency matrix W constructed in step 3 is (w ij ) s×s The specific implementation formula is as follows:

[0020]

[0021] In the formula, σ is usually taken as 1, and d ij To improve the spectral Jaccard similarity distance and reduce runtime, an improved Jaccard similarity matrix d = (d... ij ) s×s The specific implementation formula is as follows:

[0022]

[0023] In the formula, J∈R s×1 The similarity vector for each pixel is calculated using the following formula:

[0024] J i =b 00 +b 11

[0025] In the formula b 00 Let b be the number of spectral curves of the i-th dictionary atom whose first derivative is less than or equal to 0. 11 Let be the number of spectral curves of the i-th dictionary atom whose first derivative is greater than 0.

[0026] Furthermore, in step 3, a graph regularization constraint is added to the collaborative representation to construct a graph-constrained collaborative representation model, the specific functional form of which is:

[0027]

[0028] Where Γ y L represents the spectral distance between the pixel to be measured and the dictionary atom. sym The specific formula for calculating the symmetric normalized Laplace matrix is ​​as follows:

[0029]

[0030] In the formula, Let W be the degree matrix of the adjacency matrix, and L be the Laplace matrix of the adjacency matrix. The closed-form solution of the coefficient vector is specifically expressed as follows:

[0031]

[0032] Furthermore, the global anomaly significance weights constructed in step 3 using the optimal clustering framework and the spectral residual model are specifically implemented using the following formula:

[0033]

[0034] In the formula For the inverse Fourier transform, P(f) is the phase spectrum of the logarithmic spectrum L(f), and R(f) is the spectral residual. The specific calculation formula is as follows:

[0035]

[0036]

[0037] Where h n (f) is a local averaging filter. This is the average logarithmic amplitude spectrum.

[0038] Furthermore, the specific implementation formula for the nonlinear fusion method in step 4 is as follows:

[0039]

[0040] In the formula, S represents the initial anomaly detection result of the graph-constrained collaborative representation model, where S is the global anomaly saliency weight.

[0041] Working principle: The similarity in the original space is transferred to the feature space to preserve the local geometric structure in the dictionary atomic space. Cooperative and graph regularization terms are combined to constrain the coefficient matrix for better reconstruction. Simultaneously, an improved Jaccard similarity matrix and spectral distance are used to construct the adjacency matrix to enhance the effect of graph regularization constraints. Furthermore, utilizing the global information of hyperspectral imagery, an optimal clustering band selection method is adopted to determine the three bands with the richest information. A spectral residual model is used to construct global anomaly saliency weights for the selected bands. Finally, the established graph-constrained cooperative representation model is nonlinearly fused with the constructed global anomaly saliency weights. The method implemented in this invention not only improves the accuracy and stability of anomaly detection under different target scenarios and has strong applicability, but also has important theoretical and practical significance for advancing research related to anomaly detection in hyperspectral remote sensing images.

[0042] This invention proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the global anomaly weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method.

[0043] This invention proposes a computer program product, including a computer program and / or instructions, which, when executed by a processor, implements the aforementioned globally anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method.

[0044] Beneficial Effects: The globally anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method described in this invention has the following beneficial effects compared to existing technologies:

[0045] 1. This invention improves the reconstruction effect of the pixels under test by adding graph regularization constraints to the cooperative representation and constructing a graph-constrained cooperative representation model constraint coefficient matrix.

[0046] 2. This invention utilizes the global features of hyperspectral images, introducing information theory into the field of anomaly detection. Redundant parts are treated as background, and variable parts as anomalies. An optimal clustering framework and spectral residual model are used to construct a global anomaly saliency weight to highlight anomalies. Finally, the established graph-constrained collaborative representation model and the constructed global anomaly saliency weight are nonlinearly fused to obtain the final result. It has good detection performance under different data scenarios and strong robustness and applicability.

[0047] 3. To improve the constraint effect of the graph regularization term on the coefficient matrix, a Jaccard similarity coefficient is added to the spectral distance to construct the adjacency matrix. In addition, a matrix calculation method is used to construct the Jaccard similarity matrix between dictionary atoms, which reduces the running time. Attached Figure Description

[0048] Figure 1 This is a flowchart of the globally anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method described in this invention;

[0049] Figure 2 (a) and (b) in the example are the hyperspectral remote sensing image and the anomalous pixel distribution map, respectively.

[0050] Figure 3 The surface plot for searching the optimal regularization parameters λ and β using the method of this invention;

[0051] Figure 4 In the diagrams, (a) to (e) are the anomaly detection results obtained by algorithms RX, CRD, GTVLRR, VABS, and GS-GCR, respectively.

[0052] Figure 5 ROC curves for five anomaly detection methods: RX, CRD, GTVLRR, VABS, and GS-GCR. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0054] Example 1

[0055] The experimental data for Los Angeles was collected by the AVIRIS sensor in the Los Angeles Airport area of ​​California, USA. It contains 100×100 pixels, 205 effective bands, and a spatial resolution of 7.1m. The corresponding hyperspectral remote sensing image and anomalous pixel distribution map are shown below. Figure 2 As shown in (a) and (b) in the figure. The background of this dataset is relatively complex, mainly consisting of hangars, airport runways, and aprons. The anomalies are the two central aircraft, consisting of 87 pixels each. In the experiment, the outer window size was set to 19, the inner window size was set to 17, and the parameter c was set to 0.5.

[0056] The anomaly detection method based on globally weighted hyperspectral remote sensing images with constrained collaborative representation, as described in this invention, has the following process: Figure 1 As shown, the specific implementation steps are as follows:

[0057] (1) Perform data preprocessing on the original hyperspectral remote sensing image data and convert it into a two-dimensional matrix X of size b×n, where b is 205 and n is 10000;

[0058] (2) For each pixel, obtain the dictionary atom set X using the sliding double window model. s ∈R b×s , where s is the number of atoms in the dictionary;

[0059] (3) Calculate dictionary atom x i and x j Weighted distance w between ij Obtain the adjacency matrix W = (w ij ) s×s The symmetric normalized Laplace matrix L is obtained using the adjacency matrix W. sym ;

[0060] (4) Using the normalized Laplace matrix L sym Obtain the graph constraint regularization term, add the graph constraint regularization term to the collaborative representation model, and construct the graph constraint collaborative representation model;

[0061] (5) Based on the objective function constructed by the graph-constrained collaborative representation model, the closed-form solution of the coefficient vector α is obtained to obtain the initial anomaly detection results.

[0062] (6) Use the Optimal Clustering Framework (OCF) method to select bands for hyperspectral remote sensing images, obtain the image I(x) generated by the three bands with the richest information, and construct a grayscale image of image I(x).

[0063] (7) Perform Fourier transform on the grayscale image, and perform logarithmic transform on the grayscale image A(f) after Fourier transform to obtain the logarithmic amplitude spectrum L(f).

[0064] (8) Convolve the logarithmic spectrum L(f) using a local average filter to obtain the average logarithmic amplitude spectrum L(f) and the phase spectrum P(f) of the logarithmic spectrum L(f).

[0065] (9) Obtain the spectral residual R(f) using the spectral residual model. Specifically, subtract the average logarithmic amplitude spectrum L(f) from the logarithmic spectrum L(f).

[0066] (10) Perform inverse Fourier transform on the spectral residual R(f) and smooth it using a Gaussian filter G(x) to obtain the global anomaly significance weights.

[0067] (11) Using nonlinear fusion formula The initial anomaly detection results The final anomaly detection result is obtained by fusing it with the global anomaly significance weight; where e is the final anomaly detection result, c is the adjustment parameter of the constant term, and S represents the global anomaly significance weight.

[0068] (12) Output the final anomaly detection result image.

[0069] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described globally anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method.

[0070] In one embodiment, a computer program product is provided, including a computer program / instruction that, when executed by a processor, implements the globally anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method.

[0071] Anomaly detection method (GS-GCR) of this invention was experimentally analyzed on hyperspectral remote sensing image data, and compared with some representative anomaly detection algorithms, mainly including the classic RX algorithm, as well as CRD, GTVLRR, and VABS algorithms. First, the settings of the regularization parameters λ and β in the GS-GCR algorithm of this invention were analyzed. The settings of parameters λ and β are both within the range of {1e...}-5 ,1e -4 ,1e -3 ,1e -2 ,1e -1 ,1e 0 ,1e 1 Between these points, as the parameters change, the area under the ROC curve (AUC) for anomaly detection changes as follows: Figure 3 As shown, in this example, parameters λ and β are set to 1e. 1 and 1e -1 . Figure 4 The anomaly detection results of various algorithms are shown, including five algorithms: RX, CRD, GTVLRR, VABS, and GS-GCR. The figures demonstrate that the method presented in this invention (GS-GCR) exhibits better background suppression and more pronounced detection of anomalous targets. Furthermore, the ROC curves of each anomaly detection algorithm on the Los Angeles dataset are shown below. Figure 5 As shown in Table 1, the corresponding area under the ROC curve (AUC) indicates that GS-GCR achieves better anomaly detection results. In summary, the method proposed in this invention exhibits superior performance compared to other similar anomaly detection algorithms.

[0072] Table 1 Figure 5 Area under the ROC curve (%) for each algorithm

[0073]

Claims

1. A globally weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method, characterized in that: Includes the following steps: Step 1: Convert the hyperspectral remote sensing image into a two-dimensional matrix. Where b is the number of bands and n is the number of pixels; Step 2: For each pixel, construct a local dictionary set using a sliding double window. ,in The number of atoms in the dictionary; Step 3: Calculate the adjacency matrix between dictionary atoms. According to the adjacency matrix Obtain the symmetric normalized Laplace matrix The obtained symmetric normalized Laplace matrix is ​​used to construct a graph regularization term. The graph regularization term is used as one of the constraints to add a constraint term to the optimized collaborative representation model. The graph-constrained collaborative representation model is constructed, and the closed-form solution of the coefficient vector is obtained to obtain the initial anomaly detection results. The global anomaly significance weights of hyperspectral images are calculated using global information and spectral residual models. Step 4: Nonlinearly fuse the initial anomaly detection results and the global anomaly significance weights; Step 5: Output the fused anomaly detection results; The method for calculating the global anomaly significance weights of hyperspectral images using global information and spectral residual models includes the following steps: (a) Using the optimal clustering framework method to select bands in hyperspectral remote sensing images, the images generated from the three bands with the richest information are obtained. Then, the grayscale image of the image is obtained; (b) Perform a Fourier transform on the grayscale image, and convert the Fourier-transformed grayscale image... Perform a logarithmic transformation to obtain the logarithmic amplitude spectrum. ; (c) Using a local averaging filter to analyze the logarithmic spectrum Perform a convolution operation to obtain the average logarithmic amplitude spectrum. ; (d) Subtracting the mean logarithmic amplitude spectrum from the logarithmic amplitude spectrum yields the spectral residual. ; (e) Spectral residuals Perform inverse Fourier transform and smooth using a Gaussian filter to obtain global anomaly significance weights; The adjacency matrix The calculation formula is: In the formula, For dictionary atoms and Weighted distance between them It is a constant. For improved spectral Jaccard similarity distance.

2. The anomaly detection method based on globally weighted hyperspectral remote sensing images with constrained collaborative representation according to claim 1, characterized in that: Improved spectral Jaccard similarity distance matrix The calculation formula is: In the formula, I is the identity matrix. For the number of bands, The similarity vector for each pixel is calculated using the following formula: In the formula Let be the number of spectral curves of the i-th dictionary atom whose first derivative is less than or equal to 0. Let be the number of spectral curves of the i-th dictionary atom whose first derivative is greater than 0.

3. The anomaly detection method based on globally weighted hyperspectral remote sensing images with constrained collaborative representation according to claim 1, characterized in that: Based on the collaborative representation, a graph-constrained collaborative representation model is constructed by adding a graph regularization constraint, and the calculation formula is: in, Represents the spectral vector of the pixel to be measured. Represents the encoding coefficient vector. and For regularization parameters, The spectral distance between the pixel to be measured and the dictionary atom is... Let Tr(·) be the symmetric normalized Laplace matrix, and let Tr(·) denote the Tr function.

4. The anomaly detection method based on globally weighted hyperspectral remote sensing images with constrained collaborative representation according to claim 1, characterized in that: The global anomaly significance weight is calculated using the following formula: In the formula, S(x) is the global anomaly significance weight, and G(x) is the Gaussian filter function. This is the inverse Fourier transform. Logarithmic spectrum phase spectrum, This represents the spectral residual.

5. The anomaly detection method based on globally weighted hyperspectral remote sensing images with constrained collaborative representation according to claim 4, characterized in that: The formula for calculating the spectral residual is: in For local averaging filters, This is the average logarithmic amplitude spectrum.

6. The anomaly detection method based on globally weighted hyperspectral remote sensing images with constrained collaborative representation according to claim 1, characterized in that: The initial anomaly detection results and the global anomaly significance weights are nonlinearly fused, and the calculation formula is as follows: In the formula, This represents the initial anomaly detection results for the graph-constrained collaborative representation model. The weight for global anomaly significance. The results of the fused anomaly detection are as follows. This indicates the adjustment parameter, which is a constant term.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the globally anomaly-weighted hyperspectral remote sensing image constrained collaborative representation anomaly detection method as described in any one of claims 1 to 6.

8. A computer program product, comprising a computer program and / or instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the globally anomaly-weighted hyperspectral remote sensing image constraint-coordinated representation anomaly detection method as described in any one of claims 1 to 6.

Images