A hyperspectral anomaly detection method based on background reconstruction subtraction of multi-feature combination

Abstract

The present application relates to a kind of hyperspectral image anomaly detection method, the present application is based on the background reconstruction subtraction of multiple feature joint to the hyperspectral image is carried out anomaly detection.Specific method includes: first using spatial feature and spectral feature joint improved double window three-edge filtering method to reconstruct background to image;Second, for the distribution characteristics of the abnormal target in the analysis image, on the basis of traditional saliency detection method, propose the saliency feature extraction method based on global context perception to extract the saliency feature map of image;Then the saliency feature map of image and the square difference of three-edge filtering after reconstructing background map are obtained Abnormal target initial detection map;Finally, using spectral correlation coefficient to obtain the spectral weight map of image, and it is fused with initial anomaly detection map to obtain the final anomaly detection result.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral anomaly detection technology, and specifically relates to a hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features. Background Technology

[0002] Hyperspectral images possess a unique spectral integration characteristic, containing both spatial geometric information reflecting the distribution of ground features and spectral information reflecting their characteristics. Hyperspectral images can perform more in-depth detection and identification based on subtle spectral differences between ground features, an advantage unmatched by traditional optical images. Anomaly detection, as an unsupervised target detection method, requires no prior spectral information and can detect anomalous points with spectral curves significantly different from surrounding pixels. Currently, hyperspectral image anomaly detection is widely used in agricultural pest control, mineral exploration, target reconnaissance, and battlefield counterfeiting detection.

[0003] The presence of significant noise and anomalous targets in hyperspectral images interferes with background reconstruction, limiting the performance improvement of detection methods. Furthermore, analysis reveals that hyperspectral anomalous targets possess salient characteristics. Traditional anomaly detection methods, relying solely on the spatial or spectral features of hyperspectral images, suffer from limited detection performance. Summary of the Invention

[0004] The main objective of this invention is to provide a hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features, so as to improve the detection performance of hyperspectral image anomaly detection.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features includes the following steps:

[0007] Step 1, convert the original hyperspectral image H∈R M×N×B After dimensionality reduction and weighted fusion, image R is obtained. pca ∈R M×N This serves as the input image for subsequent improvements to trilateral filtering and salient feature extraction.

[0008] Step 2: In the trilateral filtering, a weighting factor representing spectral features is introduced to adjust the original hyperspectral image H and the image R. pca Processing is performed to obtain the background reconstruction image R. bg ∈R M×N ;

[0009] Step 3: Analyze the image R using an improved global context-aware saliency detection method. pca Processing is performed to obtain the saliency feature map E∈R M×N ;

[0010] Step 4: Compare the saliency feature map E with the background reconstruction map R. bg Perform subtraction and squaring operations to highlight anomalous targets and obtain an initial anomaly detection map X∈R. M×N ;

[0011] Step 5: Calculate the weight map Y∈R on the original hyperspectral image H. M×N ;

[0012] Step 6: Fuse the weight map Y with the initial anomaly detection map X to obtain the anomaly detection result G∈R. M×N .

[0013] The presence of numerous noise and outliers in hyperspectral images can interfere with the reconstructed background, limiting the performance improvement of detection methods. This invention implements a hyperspectral image anomaly detection method based on multi-feature joint background reconstruction subtraction. First, a background reconstruction map is obtained using a trilateral filtering method improved with three weighting factors: spatial Euclidean distance, pixel gray-level difference, and spectral distance. Simultaneously, a salient feature map is obtained from the dimensionality-reduced image using a global context-aware salient feature extraction method. Then, the saliency feature map is subtracted from the trilaterally filtered reconstructed background map to obtain an initial anomaly detection map. This process suppresses the background while preventing the anomaly from being suppressed, making the target more prominent. Finally, the spectral correlation coefficient is used to obtain the spectral weight map of the image, which is then fused with the initial anomaly detection map to obtain the final anomaly detection result, further improving the detection probability of the algorithm. This algorithm has advantages such as prominent target detection, good background suppression, and low false alarm rate, exhibiting excellent detection performance. Attached Figure Description

[0014] Figure 1 These are examples of the present invention. (a) is the original image in pseudo-color, and (b) is an anomaly distribution map.

[0015] Figure 2 This is a flowchart of an example of the present invention.

[0016] Figure 3 Background reconstruction image to improve the trilateral filtering method.

[0017] Figure 4 This is a salient feature extraction map of a hyperspectral image.

[0018] Figure 5 These are images showing the anomaly detection results in an example of the present invention. (a) is a pseudo-color image of the RPCA-RX detection results, and (b) is a pseudo-color image of the detection results of the present invention.

[0019] Figure 6These are three-dimensional curves of the anomaly detection results in an example of the present invention. (a) is a three-dimensional kurtosis plot of the RPCA-RX detection results, and (b) is a three-dimensional kurtosis plot of the detection results of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] This invention combines spatial features, spectral features, and saliency features of hyperspectral images to improve the robustness and other detection performance of the method. To demonstrate the effectiveness of the method, real hyperspectral images were used for anomaly detection. The experiment used a 100×100 sub-image with 189 bands from a scene image of San Diego airport acquired by an airborne visible / infrared imaging spectrometer (AVIRIS) sensor for anomaly detection, identifying three aircraft within this sub-image as anomalous targets. (Reference) Figure 1 As shown, (a) and (b) show the pseudo-color image of the image and the corresponding true distribution image of the anomalous targets, respectively.

[0022] For the SanDiego_Airport_1 image data, the main process of the hyperspectral anomaly detection method based on multi-feature joint background reconstruction subtraction of this invention is as follows: Figure 2 As shown, the process includes the following:

[0023] Step 1, convert the original hyperspectral image H∈R M×N×B After dimensionality reduction and weighted fusion, image R is obtained. pca ∈R M×N This serves as the input image for subsequent improvements to trilateral filtering and salient feature extraction.

[0024] In this step, principal component analysis is used to reduce dimensionality. The obtained principal component components are arranged in descending order of information content, and a weighted average fusion operation is performed on the first m images based on the magnitude of the eigenvalues. The specific method is as follows:

[0025] Step 1.1: Dimensionality reduction is achieved using principal component analysis. The principle is to multiply the high-dimensional vector W by the transpose of its projection vector K in an orthogonal coordinate system. The mathematical expression is:

[0026] P=K T W

[0027] In the formula, P represents the projected data. The variance S of the projected data is...

[0028] S = E(PP) T )=E[(KT W)(K T W) T ] = K T E(WW T )K

[0029] The solution for projective orthogonal bases in principal component analysis can be expressed as follows:

[0030] max S = E(PP) T ) = K T E(WW T )K stK T K = I

[0031] In the formula, I is the identity matrix. Solving the above formula using the Lagrange operator yields...

[0032] L(K)=K T E(WW T )K+λ(IK T K)

[0033] In the formula, λ is a parameter introduced by the Lagrange multiplier method. Taking the derivative of the above formula, the result is:

[0034] CK=λK

[0035] In the formula, C is E(WW) T ), where the projection vector K is the eigenvector corresponding to the eigenvalue, λ is the eigenvalue of C, and E() represents the variance operation on the vector.

[0036] Step 1.2: Obtain the m largest eigenvalues ​​from the eigenvalue matrix, which are {λ1, λ2, ..., λ...} m}, the corresponding feature vectors {K1,K2,…,K} m A new projection matrix is ​​constructed, and the resulting principal components are arranged in descending order of information content. The m principal components can be represented as {P1, P2, ..., P}. m In subsequent spatial feature extraction operations, it is only necessary to perform a weighted average fusion operation on the m images based on the feature values. The mathematical expression is:

[0037]

[0038] In the formula, R pca This is the weighted fused image, which can replace the original image for subsequent processing.

[0039] Step 2: Improve the trilateral filtering method by introducing weighting factors representing spectral features into the traditional trilateral filtering, and then applying these weighting factors to the original hyperspectral image H and image R. pca Processing is performed to obtain the background reconstruction image R. bg ∈R M×N .

[0040] Specifically, this step obtains the background reconstruction map R. bg The method can be described as follows:

[0041] Step 2.1: Introduce weighting factors representing spectral characteristics into the trilateral filter. The kernel of the improved trilateral filter is:

[0042] w tf (x)=w d (x,ε i )w g (x,ε i )w sf (x,ε i )

[0043] In the formula, x represents the target pixel, and ε i Represents the neighboring pixels of the target pixel; the first weight w d (x,ε i The second weight, w, is calculated from the Euclidean distance between two pixels in the image. g (x,ε i The third weight w is calculated from the pixel difference between two pixels. sf (x,ε i It is calculated from the spectral feature distance between two pixels.

[0044] Step 2.2: When reconstructing the center pixel, pixels that are closer to the center pixel in Euclidean distance contribute more to the reconstruction. The Euclidean distance between two pixels is:

[0045]

[0046] In the formula, σ d Here, (k,l) represents the distance weighting parameter, (u,v) represents the target pixel coordinates, and (u,v) represents the neighboring pixel coordinates.

[0047] Step 2.3: The smaller the pixel difference between a pixel and the center pixel in the image, the greater the contribution of the pixel to the reconstruction of the center pixel. The formula for calculating the pixel difference is:

[0048]

[0049] In the formula, σ g Let f(k,l) be the standard deviation of the Gaussian function, f(k,l) be the gray value of the target pixel, and f(u,v) be the gray value of the neighboring pixels.

[0050] Step 2.4, the weighting factor for the spectral characteristic distance is expressed as:

[0051]

[0052] In the formula, σsf d is the parameter output for spectral distance weighting. sf (x,ε i () represents the spectral distance between two pixels, and its mathematical expression is:

[0053]

[0054] In the formula, t is the number of bands, x r Let ε be an element of the target pixel in the r-th band. r This represents the element in the r-th band of the neighboring pixels of the target pixel, where num is the adaptive weighting coefficient, expressed as:

[0055]

[0056] Step 2.5, the improved trilateral filtering method based on spatial distance information, gray-level difference information, and spectral vector difference, can be expressed as:

[0057]

[0058] In the formula, R bg (x) represents the background reconstruction graph R. bg The value of the median pixel, i.e., the result of the target pixel x after improved trilateral filtering, is O(ε). i Let represent the input image of the trilateral filter, n be the total number of pixels in the neighborhood of the target pixel, and o(x) be the regularization parameter, which can be expressed as:

[0059]

[0060] Step 2.6: During background reconstruction, a dual-window model is used for calculation. The contribution weight of all pixels within the inner window's protected area to the reconstruction of the central pixel is zero. By sliding the dual windows, each pixel on the image is reconstructed to obtain the reconstructed background image R. bg .

[0061] In this embodiment, there are six main parameters that need to be determined: the distance weight factor s in the image saliency detection process, and the parameter σ that controls the contribution of each weight in the improved trilateral filtering method. d σ r σ sf and the inner and outer windows of the filter template w i w o Size. Let's tentatively define σ. d σ r σ sf They are {1,1,1}, w i w oThe AUC values ​​were calculated for values ​​of s, 7 and 9, with values ​​of s taking the range {1, 5, 10, 15, 20}. When s was set to 15, the AUC value reached its maximum of 0.9898, indicating relatively good detection performance. Therefore, s was set to 15 in subsequent experiments. The parameter σ controlling the contribution of each weight in the three-sided filtering method was then determined. d σ r σ sf The value of s is initially set to 15, and w is tentatively set to... i w o With values ​​of 7 and 9, the weight parameters of the trilateral filter are set to 1, 3, and 5 respectively, and the AUC value is calculated. When the weight parameters of the trilateral filter σ are... d σ r σ sf The method in this chapter achieves the highest AUC value when the weights are 5, 3, and 5 respectively; therefore, the weight parameters for the trilateral filter are set to 5, 3, and 5. The inner window size w... i Set to {5,7,9,11,13,15,17,19,21,23} respectively, with an outer window size w. o Given values ​​{7, 9, 11, 13, 15, 17, 19, 21, 23, 25}, calculate the AUC value. When w... i w o When the values ​​are set to 15 and 25 respectively, the AUC values ​​are relatively large. Therefore, to fully utilize the performance of the method, the distance weight factor s in the significance value calculation in this embodiment is set to 15, and the trilateral filter weight parameter σ is... d σ r σ sf Set to 5, 3, and 5 respectively, for inner and outer windows w i w o Sizes are 15 and 25. Figure 3 Background reconstruction image to improve the trilateral filtering method.

[0062] Step 3: Analyze the image R using an improved global context-aware saliency detection method. pca Processing is performed to obtain the saliency feature map E∈R M×N .

[0063] This step introduces a global statistical approach, highlighting the object of interest while also considering contextual elements in the background. It identifies regions of interest across the entire image, suppressing most background information while retaining information deviating from normal values. This reduces the false alarm rate and effectively extracts regions of interest from the image. The specific steps include:

[0064] Step 3.1, reduce the dimension of the image R. pca As the image to be detected, the significant difference between two pixels in the image can be expressed as:

[0065]

[0066]

[0067]

[0068] In the formula, d c (X i ,X j ) is the pixel X i and X j In Lab color space, L i and L j They represent pixels X and X respectively. i With X j brightness, a i and b i and a j and b j They represent pixels X and X respectively. i With X j Color opposite dimension, d p (X i ,X j ) represents X i and X j The Euclidean distance between them, d(X) i ,X j ) represents X i With X j The significant difference between them, where s is the weighting parameter.

[0069] Step 3.2, Image R pca The size is M×N, and the pixel X i The pixel X is obtained by summing the significance difference values ​​of all other pixels and taking the average. i The average significant difference value is expressed as:

[0070]

[0071] In the formula, S(X) i ) represents pixel X i The average significance value of the difference.

[0072] Step 3.3: Set a threshold Th (in this embodiment, the threshold Th is set to 0.02). Pixels with an average significance difference value greater than Th are considered abnormal target pixels, and pixels with a value less than Th are considered background pixels. Then, a weighted calculation is performed based on the Euclidean distance between the background pixel and its nearest abnormal target pixel. The final significance value is:

[0073]

[0074] In the formula, For pixels X i The final significant eigenvalue, d near (X i ) is the pixel X i The Euclidean distance between the pixel and the nearest anomalous target pixel, and the salient feature values ​​of all pixels together form the salient feature map E. Specifically, the final salient feature values ​​of each non-anomalous pixel are updated. That is, the saliency feature map E is obtained, such as Figure 4 As shown.

[0075] Step 4: Compare the saliency feature map E with the background reconstruction map R. bg Perform subtraction and squaring operations to highlight anomalous targets and obtain an initial anomaly detection map X∈R. M×N .

[0076] Since the final saliency feature map E contains a small amount of noise, this step compares the saliency feature map E with the background reconstruction map R obtained after trilateral filtering. bg Subtraction can suppress the background. The initial anomaly detection map is represented as follows:

[0077]

[0078] In the formula, E p and Represent the saliency feature map E and the background reconstruction map R, respectively. bg The pixel value of the p-th pixel, X p This represents the pixel value of the p-th pixel in the initial anomaly detection map.

[0079] Step 5: Calculate the weight map Y∈R on the original hyperspectral image H. M×N .

[0080] Since the initial anomaly detection map may still contain some noise, in order to further highlight the anomalous targets, the spectral weight map Y of the original hyperspectral image H is calculated using the spectral correlation coefficient. The specific steps are as follows:

[0081] For the original hyperspectral image H, the weighted graph Y calculated using the spectral correlation coefficient is represented as follows:

[0082]

[0083] In the formula, H p Let x be the weight of the p-th pixel in the original hyperspectral image. ip This represents the grayscale value of the p pixels in the i-th band. γ represents the spectral average value of the spectral curve containing the p-th pixel. i This represents the average spectral value of all pixels in the i-th band. The average spectral value of the average spectral curve γ is represented by the weight of each pixel. After obtaining the weight of each pixel, all weight results are combined into a weighted map Y.

[0084] Step 6: Fuse the weight map Y with the initial anomaly detection map X to obtain the anomaly detection result G∈R. M×N .

[0085] The fusion process in this step can be represented as follows:

[0086] G p =Y p ·X p p = 1, 2, ..., M × N

[0087] In the formula, "·" indicates that the p-th pixel in the weight map Y is multiplied by the p-th pixel in the initial detection map X, and the result G is... p Let p be the value of the p-th pixel in the anomaly detection result G, such as Figure 5 As shown.

[0088] The numerous noises and outliers in hyperspectral images can interfere with the reconstructed background, limiting the performance improvement of detection methods. This invention implements a hyperspectral image anomaly detection method based on multi-feature joint background reconstruction subtraction. First, a background reconstruction map is obtained using a trilateral filtering method improved with three weighting factors: spatial Euclidean distance, pixel gray-level difference, and spectral distance. Simultaneously, a salient feature map is obtained from the dimensionality-reduced image using a global context-aware salient feature extraction method. Then, the saliency feature map of the image is subtracted from the reconstructed background map after trilateral filtering, and the squared result is obtained to obtain an initial anomaly detection map. Finally, the spectral correlation coefficient is used to obtain the spectral weight map of the image, and this weight map is fused with the initial anomaly detection map to obtain the final anomaly detection result.

[0089] Figure 5 The results of anomaly detection using the RPCA_RX method and the method of this invention show that the RPCA_RX method does not detect abnormal targets clearly and there is a lot of residual background, while the method of this invention detects targets clearly. Figure 6 The three-dimensional mesh diagrams showing the detection results of the RPCA_RX method and the method of the present invention are shown. It can be seen that the RPCA_RX method can detect numerical peaks at the locations corresponding to the abnormal targets, but the targets are not very prominent. The method of the present invention has a better background suppression effect and can better detect abnormal target information in hyperspectral images.

Claims

1. A hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features, characterized in that, The steps include the following: Step 1, convert the original hyperspectral image H∈R M×N×B After dimensionality reduction and weighted fusion, image R is obtained. pca ∈R M×N This serves as the input image for subsequent improvements to trilateral filtering and salient feature extraction. Step 2: In the trilateral filtering, a weighting factor representing spectral features is introduced to adjust the original hyperspectral image H and the image R. pca Processing is performed to obtain the background reconstruction image R. bg ∈R M×N ; Step 3: Analyze the image R using an improved global context-aware saliency detection method. pca Processing is performed to obtain the saliency feature map E∈R M×N ; Step 4: Compare the saliency feature map E with the background reconstruction map R. bg Perform subtraction and squaring operations to highlight anomalous targets and obtain an initial anomaly detection map X∈R. M×N ; Step 5: Calculate the weight map Y∈R on the original hyperspectral image H. M×N ; Step 6: Fuse the weight map Y with the initial anomaly detection map X to obtain the anomaly detection result G∈R. M×N .

2. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 1, characterized in that, In step 1, dimensionality reduction is performed using principal component analysis (PCA). The obtained principal component components are then sorted in descending order of information content, and further sorted according to eigenvalue magnitude. m A weighted average fusion operation is performed on the images.

3. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 2, characterized in that, The principle of the principal component analysis dimensionality reduction method is to reduce the dimensionality of high-dimensional vectors. W Its projection vector in the orthogonal coordinate system K The product of transposes is expressed mathematically as: In the formula, P represents the projected data, and the variance of the projected data is... S Size: The solution for projective orthogonal bases in principal component analysis is expressed as: In the formula, I is the identity matrix. Solving the above formula using the Lagrange operator yields: In the formula, λ Introducing parameters into the Lagrange multiplier method and differentiating the above equation, the result is: In the formula, C is E ( WW T ), projection vector K The eigenvectors corresponding to the eigenvalues, λ Let C be an eigenvalue. E () represents the operation of calculating the variance of a vector; Obtained from the eigenvalue matrix m The largest eigenvalues ​​are { λ 1, λ 2, …, λ m }, the corresponding feature vector { K 1, K 2, …, K m } to form a new projection matrix, and arrange the obtained principal component components in descending order of information content, where m The principal components are represented as {P1, P2, ..., P...} m }; The mathematical expression for the weighted average fusion operation is: 。 4. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 1, characterized in that, In step 2, the background reconstruction image R is obtained. bg The steps are as follows: Step 2.1: Introduce weighting factors representing spectral characteristics into the trilateral filter. The kernel of the improved trilateral filter is: In the formula, x Indicates the target pixel. ε i Represents the neighboring pixels of the target pixel; the first Weight The second weight is calculated from the Euclidean distance between two pixels in the image. The third weight is calculated from the pixel difference between two pixels. Calculated from the spectral feature distance between two pixels; Step 2.2, the improved trilateral filtering method based on spatial distance information, gray-level difference information, and spectral vector difference, is expressed as: In the formula, Represents the background reconstruction graph R bg The value of the middle pixel, i.e., the target pixel. x The result after improved trilateral filtering, This represents the input image after trilateral filtering. n The total number of pixels in the neighborhood of the target pixel. o ( x ) is the regularization parameter, expressed as: Step 2.3: During background reconstruction, a dual-window model is used for calculation. All pixels within the inner window's protected area contribute zero weight to the reconstruction of the central pixel. By sliding the dual windows, each pixel in the image is reconstructed to obtain the reconstructed background image R. bg .

5. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 4, characterized in that, In step 2.1: When reconstructing the center pixel, pixels that are closer to the center pixel in Euclidean distance contribute more to the reconstruction. The weighting factor for the Euclidean distance between two pixels is: In the formula, σ d For distance weight parameters, ( k , l ) represents the target pixel coordinates, ( u , v () represents the coordinates of the neighboring pixels; In an image, the smaller the pixel difference from the center pixel, the greater the contribution of the pixel to the reconstruction of the center pixel. The weighting factor for the pixel difference is calculated as follows: In the formula, σ g Let be the standard deviation of the Gaussian function. f ( k , l () represents the grayscale value of the target pixel. f(u, v) The grayscale value of the neighboring pixels; The weighting factor for spectral characteristic distance is expressed as: In the formula, σ sf The parameters output are for spectral distance weighting. The mathematical expression for the spectral distance between two pixels is: In the formula, t For the number of bands, x r Represented as the target pixel at the th r Elements of each band, ε r Indicates the number of neighboring pixels of the target pixel at the th r Elements of each band, num For adaptive weighting coefficients, denoted as 。 6. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 1, characterized in that, In step 3, the steps to obtain the saliency feature map E are as follows: Step 3.1, reduce the dimension of the image R. pca As the image to be detected, the significant difference value between two pixels in the image is represented as: In the formula, d c (X i X j ) is the pixel X i and X j Color distance in Lab space, L i and L j They represent pixels X and X respectively. i With X j brightness, a i and b i as well as a j and b j They represent pixels X and X respectively. i With X j Color opposite dimensions, d p (X i X j ) represents X i and X j The Euclidean distance between them d (X i ,X j ) represents X i With X j The significant difference between them s These are weight parameters; Step 3.2, Image R pca The size is M × N , will pixel X i The significance difference values ​​of pixel X are summed with those of all other pixels and then averaged to obtain pixel X. i The average significant difference value is expressed as: In the formula, S (X i ) represents pixel X i The average significant difference value; Step 3.3, Set the threshold Th The average significance value is greater than Th Pixels smaller than 0 are considered abnormal target pixels. Th The pixels are considered background pixels; then, a weighted calculation is performed based on the Euclidean distance between the background pixels and their nearest anomalous target pixels, and the final significance value is: In the formula, For pixels X i The final significant eigenvalues, d near (X i ) is the pixel X i The Euclidean distance between the pixel and the nearest anomalous target pixel, and the salient feature values ​​of all pixels together form the salient feature map E.

7. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple feature joint methods according to claim 1, characterized in that, In step 4, the initial anomaly detection map is represented as follows: In the formula, E p and Represent the saliency feature map E and the background reconstruction map R, respectively. bg The Middle p The pixel value of X pixels. p Indicates the first anomaly detection graph in the initial anomaly detection graph. p The pixel value of each pixel.

8. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 1, characterized in that, In step 5, the weighted graph Y calculated using the spectral correlation coefficient is represented as follows: In the formula, H p The first in the original hyperspectral image p The weight of each pixel, Indicates the first On each band p grayscale value of each pixel. Indicates the first p The spectral average of the spectral curves of each pixel. Indicates the first The spectral average of all pixels in each band. The average spectral value of the average spectral curve γ is represented by the weight of each pixel. After obtaining the weight of each pixel, all weight results are combined into a weighted map Y.

9. The hyperspectral anomaly detection method based on background reconstruction subtraction using multiple features as described in claim 1, characterized in that, Step 6, the fusion process, is represented as follows: In the formula, " " indicates the first in the weighted graph Y p The pixel and the first pixel in the initial detection image X p Multiplying by 1 pixel, the result is For the anomaly detection result G, the first... p The value of each pixel.

10. A computer-readable storage medium, characterized in that, Used to store a computer program that causes a computer to perform the method as described in any one of claims 1-9.

Images