A hyperspectral anomaly detection method based on dual-branch generative adversarial network

Abstract

The present invention provides a hyperspectral anomaly detection method based on a dual-branch generative adversarial network, comprising: decomposing a low-rank rough background matrix and a sparse rough anomaly matrix from an original hyperspectral image; inputting the rough background matrix and the rough anomaly matrix into a dual-branch network model, wherein the dual-branch network model outputs a reconstructed background matrix corresponding to the rough background matrix, and the dual-branch network model outputs a reconstructed anomaly matrix corresponding to the rough anomaly matrix; fusing the reconstructed background matrix and the reconstructed anomaly matrix to obtain fused hyperspectral data; calculating a pure background mean vector and a background covariance matrix using the reconstructed background matrix; solving for anomaly response values of each spectral vector in the fused hyperspectral data based on the background mean vector and the background covariance matrix; and performing normalization processing on the anomaly response values of each spectral vector to obtain an anomaly detection result map for the entire hyperspectral image. The present invention enhances the distinguishability between background and anomaly samples and improves anomaly detection accuracy.

Description

Technical Field

[0001] The present invention relates to the field of hyperspectral image processing, and in particular to a hyperspectral anomaly detection method based on a dual-branch generative adversarial network. Background Art

[0002] Hyperspectral imagery not only captures two-dimensional spatial information about the distribution of objects, but also provides one-dimensional spectral information reflecting the inherent properties of the objects, making it of great research value. Anomaly detection in hyperspectral imagery aims to identify small targets that exhibit significant spectral differences from their surrounding backgrounds, without any prior spectral information. In practical applications, prior spectral information about the target or background is often difficult to obtain. Furthermore, traditional target detection techniques rely heavily on the performance of relevant atmospheric compensation and spectral correction algorithms. Consequently, anomaly detection in hyperspectral imagery has garnered increasing attention and has become a research hotspot.

[0003] For example, the RX algorithm proposed by Reed and Xiaoli is one of the most classic hyperspectral anomaly detection algorithms. It assumes that the background in hyperspectral images follows a multivariate Gaussian distribution model and is derived from the generalized likelihood ratio test. The RX detector operator is ultimately formulated as the Mahalanobis distance between each spectral vector in the scene and the background mean vector represented by the global spectral mean. However, in practice, due to the complexity of hyperspectral images, their backgrounds often do not follow a single multivariate Gaussian distribution model. To address this issue, a series of improved RX algorithms have been proposed. For example, the local RX algorithm uses a concentric dual window to traverse each pixel in the hyperspectral image and perform RX detection within each local region bounded by the inner and outer windows. However, local RX is clearly unable to overcome the problem of anomalous pixels contaminating the local background and is heavily dependent on the size of the dual windows. Furthermore, Carlotto et al. assume that real hyperspectral images are composed of a mixture of multiple Gaussian distribution models. They use a spectral clustering algorithm to divide the hyperspectral image into clusters and then perform RX detection within each cluster. While this approach alleviates the statistical distribution issues of real hyperspectral images to some extent, it is still subject to abnormal contamination in background estimation, and the performance of the spectral clustering algorithm has a significant impact on its actual detection results. In addition, the above traditional methods only utilize low-order data information in hyperspectral images, completely ignoring the high-order, nonlinear, and abstract inherent data characteristics contained therein.

[0004] In recent years, several deep learning-based hyperspectral anomaly detection methods have been proposed. In 2017, Li et al. developed a supervised convolutional neural network model for hyperspectral anomaly detection. This method first iteratively trains a multi-layer CNN network on a training dataset homologous to the target hyperspectral dataset, using a reference image with categorical information as training labels. This network model is then transferred to the target hyperspectral dataset and uses the differences between neighboring pixel pairs output by the network to perform anomaly detection. As a pioneer in deep learning-based hyperspectral anomaly detection, this method is highly innovative. However, it is implemented in a supervised mode and relies entirely on the availability of a reference image with categorical information, which is difficult to obtain in practice. Furthermore, the complexity of the network model makes this method computationally expensive, which limits its application to some extent. In an unsupervised mode, Bati et al. pioneered the application of an autoencoder model to hyperspectral anomaly detection to characterize backgrounds of varying complexity. This method inputs all spectral samples in the hyperspectral image into an autoencoder, which outputs the corresponding decoded samples. Finally, the reconstruction error between the original and decoded images is measured to determine anomalies. Although this method is executed in an unsupervised mode, the abnormal signals in the input samples will undoubtedly seriously interfere with the normal update of the network parameters, ultimately causing the network model to capture abnormal information while learning the background distribution. Moreover, this method completely ignores the local features and structural features between each sample in the hyperspectral image, resulting in low background-anomaly discrimination.

[0005] It can be seen that the existing hyperspectral anomaly detection methods not only have the defect of low background-anomaly distinction, but also the anomaly detection accuracy needs to be improved. Summary of the Invention

[0006] The present invention provides a hyperspectral anomaly detection method based on a dual-branch generative adversarial network, which aims to enhance the distinguishability between background and abnormal samples and improve the anomaly detection accuracy.

[0007] The technical solutions for achieving the purpose of the present invention are as follows:

[0008] A hyperspectral anomaly detection method based on a dual-branch generative adversarial network, comprising:

[0009] Step 1: Decompose the low-rank rough background matrix B from the original hyperspectral image X 2D and the sparse rough anomaly matrix A 2D ;

[0010] Step 2: The rough background matrix B 2D and the rough anomaly matrix A 2DInput the dual-branch network model, which corresponds to the rough background matrix B 2D Output reconstructed background matrix The two-branch network model corresponds to the rough anomaly matrix A 2D Output reconstructed anomaly matrix

[0011] Step 3: Fusion of the reconstructed background matrix With the reconstructed anomaly matrix Obtain the fused hyperspectral data and reconstruct the background matrix using the Calculate the clean background mean vector and background covariance matrix;

[0012] Step 4: using an RX detector operator to solve the abnormal response value of each spectral vector in the hyperspectral data according to the background mean vector and the background covariance matrix;

[0013] Step 5: Normalize the abnormal response values of each spectral vector to obtain the abnormal detection result map of the entire hyperspectral image.

[0014] The present invention utilizes the background network branch in the dual-branch network model to extract a pure, uniform, and Gaussian-distributed reconstructed background matrix from the rough background matrix, and simultaneously utilizes the abnormal network branch in the dual-branch network model to extract a pure reconstructed abnormal matrix from the rough abnormal matrix; the present invention utilizes the separation enhancement constraint of the reconstruction space to explicitly enhance the spectral discrimination between the reconstructed background matrix and the reconstructed abnormal matrix; the reconstructed background matrix and the reconstructed abnormal matrix are fused to obtain fused hyperspectral data, and the reconstructed background matrix is used to calculate a pure background mean vector and a background covariance matrix; based on the calculated background mean vector and background covariance matrix, an RX detector operator is used to solve the abnormal response value of each spectral sample in the fused hyperspectral data, thereby obtaining an abnormality detection result diagram of the entire hyperspectral image.

[0015] As a further limitation of the present invention, the step 1 is specifically:

[0016] Step (11): convert the original hyperspectral image X into a two-dimensional hyperspectral matrix X with a size of N×S. 2D , and the two-dimensional hyperspectral matrix X 2D Perform standardization to obtain the standardized two-dimensional hyperspectral matrix X 2D ; Where: N is the number of pixels in X, S is the number of bands in X;

[0017] Step (12): Using the GoDec algorithm, the normalized two-dimensional hyperspectral matrix X 2D Perform low-rank and sparse matrix decomposition to extract the low-rank rough background matrix B2D and the sparse rough anomaly matrix A 2D .

[0018] As a further limitation of the present invention, the two-dimensional hyperspectral matrix X 2D Perform standardization processing, specifically:

[0019] For a two-dimensional hyperspectral matrix X with size N×S 2D , calculate the mean μ of each column element corresponding to each band i and variance δ i , and X 2D The following normalization formula (1) is performed on each column element in:

[0020]

[0021] In formula (1), i=1,2,...,S represents the column index. After normalization, the two-dimensional hyperspectral matrix X 2D The data of each band in is converted to a value with a mean of 0 and a variance of 1.

[0022] As a further limitation of the present invention, the GoDec algorithm is used to obtain the normalized two-dimensional hyperspectral matrix X. 2D Perform low-rank and sparse matrix decomposition to extract the low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D , specifically:

[0023] (1) The mathematical model of low-rank and sparse matrix decomposition is X 2D =B 2D +A 2D +N 2D ;

[0024] (2) The specific decomposition steps of the GoDec algorithm include:

[0025] (2a) Set the rank upper bound r of the low-rank matrix, the sparsity upper bound k of the sparse matrix, and the error tolerance ε;

[0026] (2b) Initialize the low-rank and sparse matrix: Where, the superscript represents the number of iterations t=0, and 0 represents a two-dimensional matrix of size N×S with all element values 0;

[0027] (2c) If where ||·|| F Represents the F norm of the matrix, then stop the iteration and output and Execute step (2j), otherwise, proceed to the next step;

[0028] (2d) Let the iteration number \(t = t + 1\), and optimize the low-rank matrix \(B\) using the following formula (2). 2D :

[0029]

[0030] In formula (2), the power exponent \(q>0\). is the optimized low-rank matrix;

[0031] (2e) Use a random matrix to construct the bilateral random projection of the low-rank matrix : Use the commands \(randn(S,r)\) and \(randn(N,r)\) to generate a random vector of size \(S\times r\) and a random vector of size \(N\times r\), denoted as the right projection matrix \(M_1\) and the left projection matrix \(M_2\), and construct the right random projection and the left random projection

[0032] (2f) Update the left projection matrix \(M_2 = Y_1\), and update the left random projection Update the right projection matrix \(M_1 = Y_2\), and update the right random projection

[0033] (2g) Perform QR decomposition on the bilateral random projections \(Y_1\) and \(Y_2\): \(Y_1 = Q_1R_1\) and \(Y_2 = Q_2R_2\);

[0034] (2h) If \(rank(M_2 T Y_1)<r\), then let \(r = rank(M_2 T Y_1)\), and return to step (2a); otherwise, continue to the next step;

[0035] (2i) Update the matrix Update the matrix where \(P Ω (\cdot)\) represents the projection of the matrix onto the set \(\Omega\), and \(\Omega\) is the non-zero subset composed of the first \(kN\) largest elements of , and return to step (2c);

[0036] (2j) Finally, and are the low-rank rough background matrix \(B 2D and the sparse rough anomaly matrix \(A 2D .

[0037] As a further limitation of the present invention, the specific steps of step two are as follows:

[0038] Step (21), input each spectral vector in the rough background matrix \(B 2D into the background network branch, and at the same time input the rough anomaly matrix \(A 2DEach spectral vector in is input into the abnormal network branch, and the reconstructed background matrix output by the background network branch is and the reconstructed anomaly matrix output by the anomaly network branch Apply the separation enhancement constraint and iteratively train the two-branch network model 2000 times to obtain the trained two-branch network model;

[0039] Step (22): For the trained dual-branch network model, extract the output of the decoder in the background network branch to reconstruct the background matrix Extract the output of the decoder in the abnormal network branch, which is the reconstructed abnormal matrix

[0040] As a further limitation of the present invention, step three is specifically:

[0041] Step (31), reconstructing the background matrix and the reconstructed anomaly matrix Fusion is performed to obtain the fused hyperspectral image

[0042] Step (32), solve the reconstructed background matrix by formula (3) The mean vector and covariance matrix of all spectral vectors in are respectively denoted as background mean vector μ B and the background covariance matrix C B ;

[0043]

[0044] In formula (3), Represents the reconstructed background matrix The spectrum vector corresponding to the i-th row in .

[0045] As a further limitation of the present invention, step four is specifically:

[0046] According to the background mean vector μ B and the background covariance matrix C B , using the RX detector operator, the fused hyperspectral image Perform anomaly detection on each spectral vector in to obtain the abnormal response value of each spectral vector;

[0047] The RX detector operator is calculated by formula (4):

[0048]

[0049] In formula (4), The fused hyperspectral image The spectral vector corresponding to the i-th row in , It is C B The inverse matrix of , the obtained RX(i) is the abnormal response value of the i-th spectral vector, and RX is the abnormal response matrix composed of the abnormal response values of all spectral vectors.

[0050] As a further limitation of the present invention, step five is specifically:

[0051] Normalize the abnormal response matrix RX obtained in step 4 to obtain the abnormal detection result map of the entire hyperspectral image. The implementation steps are as follows:

[0052] Step (51) performs normalization processing on the abnormal response matrix RX obtained in step 4 to obtain the normalized abnormal response matrix Re, which is normalized using the following formula (5):

[0053]

[0054] In formula (5), min(RX) and max(RX) are the minimum and maximum values of all elements in the abnormal response matrix RX, respectively;

[0055] Step (52): Arrange the normalized abnormal response matrix Re according to the spatial position and convert it into a two-dimensional abnormal response matrix Re with a size of H×W. 2D , where H and W are the spatial height and spatial width of the original hyperspectral image, respectively;

[0056] Step (53), use Matlab to transform the two-dimensional abnormal response matrix Re 2D Displayed as an image, it is the anomaly detection result graph.

[0057] As a further limitation of the present invention, in step (21):

[0058] The background network branch consists of an encoder, a decoder, a Gaussian adversarial network of the potential layer, and a low-dimensional manifold constraint of the potential layer. The encoder in the background network branch receives a low-rank rough background matrix B 2D The spectral vectors in the latent layer are learned through the Gaussian adversarial network and the low-dimensional manifold constraint of the latent layer to extract the intrinsic background spectral features. Finally, the decoder outputs a pure and uniform reconstructed background matrix that satisfies the Gaussian distribution.

[0059] The abnormal network branch consists of an encoder and a decoder. The encoder in the abnormal network branch receives a sparse rough abnormal matrix A 2D For each spectral vector in , the decoder outputs a clean reconstructed anomaly matrix;

[0060] The separation enhancement constraint is calculated by the following formula:

[0061]

[0062] in is the rough background matrix B of the decoder input in the background network branch 2D The reconstruction vector output by the i-th spectral vector in, is the rough anomaly matrix A of the decoder input in the anomaly network branch 2D The reconstructed vector output by the i-th spectral vector in .

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] (1) The present invention designs an unsupervised two-branch network model. Based on the GoDec decomposition, the background network branch and the anomaly network branch are used to extract the intrinsic background spectral features and intrinsic anomaly spectral features in the hyperspectral image in parallel. Thus, the spectral features are extracted and purified through the data pre-separation operation and the feature extraction capability of the network model, overcoming the signal contamination problem commonly faced by existing methods.

[0065] (2) The present invention imposes a separation enhancement constraint term in the reconstruction space of the background network branch and the abnormal network branch, so that the dual-branch network affects the parameter update process of the other party while updating its own network parameters, directly enhancing the response discrimination between background samples and abnormal samples, making abnormal samples easier to distinguish from the background, thereby improving the accuracy of anomaly detection.

[0066] (3) The present invention extracts a low-dimensional background manifold from a low-rank rough background matrix and uses it to constrain the latent layer spectral features of the background network branch, thereby retaining the similarity relationship between local neighborhood samples in the original background data during the network training process. While utilizing spectral information, it realizes the mining of inherent local structural features, which is conducive to further improving the accuracy of anomaly detection.

[0067] (4) The present invention designs Gaussian adversarial training in the latent layer of the background network branch, so that the background network branch outputs a pure and uniform background matrix that satisfies the Gaussian distribution, and estimates the background statistics (background mean vector and background covariance matrix) used for RX detection from the output background matrix, thereby overcoming the long-standing problems of mismatch between background and Gaussian distribution and anomaly contamination faced by the RX algorithm, and improving the final anomaly detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A flow chart of a hyperspectral anomaly detection method based on a dual-branch generative adversarial network provided by the present invention;

[0069] Figure 2This is a schematic diagram of the implementation process of a hyperspectral anomaly detection method based on a dual-branch generative adversarial network of the present invention;

[0070] Figure 3 A comparison chart of anomaly detection results of various anomaly detection methods;

[0071] Figure 4 Comparison of ROC curves obtained for various anomaly detection methods;

[0072] Figure 5 This is a comparison table of AUC values obtained by various anomaly detection methods on the Urban hyperspectral dataset. DETAILED DESCRIPTION

[0073] The present invention is described in detail below with reference to the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in this field based on these embodiments are all within the scope of protection of the present invention.

[0074] Example 1:

[0075] See also Figure 1 The embodiment of the present invention provides a hyperspectral anomaly detection method based on a dual-branch generative adversarial network, comprising:

[0076] S110, decompose the low-rank rough background matrix B from the original hyperspectral image X 2D and the sparse rough anomaly matrix A 2D ;

[0077] S120, the rough background matrix B 2D and the rough anomaly matrix A 2D Input the dual-branch network model, which corresponds to the rough background matrix B 2D Output reconstructed background matrix The two-branch network model corresponds to the rough anomaly matrix A 2D Output reconstructed anomaly matrix

[0078] S130, fusion and reconstruction background matrix and reconstruct the anomaly matrix Get the fused hyperspectral data and reconstruct the background matrix Calculate the clean background mean vector and background covariance matrix;

[0079] S140, using an RX detector operator to solve the abnormal response value of each spectral vector in the fused hyperspectral data according to the background mean vector and the background covariance matrix;

[0080] S150 , performing normalization processing on the abnormal response values of each spectral vector to obtain an abnormality detection result graph of the entire hyperspectral image.

[0081] The embodiment of the present invention utilizes the background network branch in the dual-branch network model to extract a pure, uniform, and Gaussian-distributed reconstructed background matrix from the rough background matrix, and simultaneously utilizes the abnormal network branch in the dual-branch network model to extract a pure reconstructed abnormal matrix from the rough abnormal matrix; the present invention utilizes the separation enhancement constraint of the reconstruction space to explicitly enhance the spectral discrimination between the reconstructed background matrix and the reconstructed abnormal matrix; the reconstructed background matrix and the reconstructed abnormal matrix are fused to obtain fused hyperspectral data, and the reconstructed background matrix is used to calculate a pure background mean vector and a background covariance matrix; based on the calculated background mean vector and background covariance matrix, an RX detector operator is used to solve the abnormal response value of each spectral sample in the fused hyperspectral data, thereby obtaining an abnormality detection result diagram of the entire hyperspectral image.

[0082] The above S110 is specifically as follows:

[0083] S111, convert the original hyperspectral image X into a two-dimensional hyperspectral matrix X of size N×S 2D , and the two-dimensional hyperspectral matrix X 2D Perform standardization to obtain the standardized two-dimensional hyperspectral matrix X 2D ; Where: N is the number of pixels in X, S is the number of bands in X;

[0084] S112, using the GoDec algorithm, from the standardized two-dimensional hyperspectral matrix X 2D Perform low-rank and sparse matrix decomposition to extract the low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D .

[0085] The above two-dimensional hyperspectral matrix X 2D Perform standardization processing, specifically:

[0086] For a two-dimensional hyperspectral matrix X with size N×S 2D , calculate the mean μ of each column element corresponding to each band i and variance δ i , and X 2D The following normalization formula (1) is performed on each column element in:

[0087]

[0088] In formula (1), i=1,2,...,S represents the column index. After normalization, the two-dimensional hyperspectral matrix X 2DThe data of each band in is converted to a value with a mean of 0 and a variance of 1.

[0089] The above-mentioned GoDec algorithm is used to obtain the normalized two-dimensional hyperspectral matrix X 2D Perform low-rank and sparse matrix decomposition to extract the low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D , specifically:

[0090] (1) The mathematical model of low-rank and sparse matrix decomposition is X 2D =B 2D +A 2D +N 2D ;

[0091] (2) The specific decomposition steps of the GoDec algorithm include:

[0092] (2a) Set the rank upper bound r of the low-rank matrix, the sparsity upper bound k of the sparse matrix, and the error tolerance ε;

[0093] (2b) Initialize the low-rank and sparse matrix: Where, the superscript represents the number of iterations t=0, and 0 represents a two-dimensional matrix of size N×S with all element values 0;

[0094] (2c) If where ||·|| F Represents the F norm of the matrix, then stop the iteration and output and Execute step (2j), otherwise, proceed to the next step;

[0095] (2d) Let the number of iterations t = t + 1, and use the following formula (2) to optimize the low-rank matrix B 2D :

[0096]

[0097] In formula (2), the power index q>0, is the optimized low-rank matrix;

[0098] (2e) Constructing a low-rank matrix using a random matrix Bilateral random projection: Use the randn(S,r) and randn(N,r) commands to generate a random vector of size S×r and a random vector of size N×r, denoted as the right projection matrix M1 and the left projection matrix M2, and construct the right random projection and left random projection

[0099] (2f) Update the left projection matrix M2 = Y1, update the left random projection Update the right projection matrix M1 = Y2, and update the right random projection

[0100] (2g) Perform QR decomposition on the bilateral random projections Y1 and Y2: Y1 = Q1R1 and Y2 = Q2R2;

[0101] (2h) If rank(M2 T Y1) < r, then set r = rank(M2 T Y1), and return to step (2a); otherwise, continue to the next step;

[0102] (2i) Update the matrix Update the matrix where P Ω (·) represents the projection of the matrix onto the set Ω, and Ω is a non - zero subset composed of the first kN largest elements of , and return to step (2c);

[0103] (2j) Finally, and are the low - rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D .

[0104] Specifically, S120 is as follows:

[0105] S121. Input each spectral vector in the rough background matrix B 2D into the background network branch, and at the same time input each spectral vector in the rough anomaly matrix A 2D into the anomaly network branch. Apply the separation - degree enhancement constraint to the reconstructed background matrix output by the background network branch and the reconstructed anomaly matrix output by the anomaly network branch, and iteratively train the double - branch network model 2000 times to obtain the trained double - branch network model;

[0106] More specifically, the background network branch in S121 consists of an encoder, a decoder, a Gaussian adversarial network in the latent layer, and a low - dimensional manifold constraint in the latent layer, and their respective structures are as follows:

[0107] Encoder structure: input layer (number of nodes is the number of bands S in the original hyperspectral image) → hidden layer (number of nodes is 500) → hidden layer (number of nodes is 500) → output layer (i.e., the latent layer, number of nodes is 20);

[0108] Decoder structure: input layer (i.e., the latent layer, number of nodes is 20) → hidden layer (number of nodes is 500) → hidden layer (number of nodes is 500) → output layer (i.e., the reconstruction layer, number of nodes is the number of bands S in the original hyperspectral image);

[0109] The Gaussian adversarial network structure of the latent layer is: input layer (i.e. latent layer, number of nodes is 20) → hidden layer (number of nodes is 500) → hidden layer (number of nodes is 500) → output layer (number of nodes is 1);

[0110] Low-dimensional manifold constraint of the potential layer: Using the local linear embedding LLE method, the low-rank rough background matrix B decomposed by S110 is 2D Extract the low-dimensional background manifold B ML , which has a rough background matrix B with the encoder 2D The output potential data Z has the same dimension. Therefore, a low-dimensional manifold constraint is imposed on the potential layer, as follows:

[0111]

[0112] in is the low-dimensional background manifold B ML The spectral vector corresponding to the i-th row in z i is the i-th latent spectral vector in Z, and N is the number of pixels in the original hyperspectral image.

[0113] More specifically, the abnormal network branch in S121 consists of an encoder and a decoder. The encoder in the abnormal network branch receives the sparse rough abnormal matrix A 2D For each spectral vector in , the decoder outputs a clean reconstructed anomaly matrix;

[0114] The structures corresponding to each module in the abnormal network branch are as follows:

[0115] Encoder structure: input layer (the number of nodes is the number of bands S in the original hyperspectral image) → hidden layer (the number of nodes is 500) → hidden layer (the number of nodes is 500) → output layer (i.e., potential layer, the number of nodes is 20);

[0116] Decoder structure: input layer (i.e., potential layer, the number of nodes is 20) → hidden layer (the number of nodes is 500) → hidden layer (the number of nodes is 500) → output layer (i.e., reconstruction layer, the number of nodes is the number of bands S of the original hyperspectral image).

[0117] More specifically, the separation enhancement constraint in S121 is calculated using the following formula:

[0118]

[0119] in is the rough background matrix B of the decoder input in the background network branch 2D The reconstruction vector output by the i-th spectral vector in, is the rough anomaly matrix A of the decoder input in the anomaly network branch 2D The reconstructed vector output by the i-th spectral vector in .

[0120] S122. For the trained dual-branch network model, extract the output of the decoder in the background network branch, which is the reconstructed background matrix Extract the output of the decoder in the abnormal network branch, which is the reconstructed abnormal matrix

[0121] The above S130 is specifically as follows:

[0122] S131, reconstruct the background matrix and reconstruct the anomaly matrix Fusion is performed to obtain the fused hyperspectral image

[0123] S132, solve the reconstructed background matrix by formula (3) The mean vector and covariance matrix of all spectral vectors in are respectively denoted as background mean vector μ B and the background covariance matrix C B ;

[0124]

[0125] In formula (3), Represents the reconstructed background matrix The spectral vector corresponding to the i-th row in .

[0126] The above S140 is specifically:

[0127] According to the background mean vector μ B and the background covariance matrix C B , using the RX detector operator, the fused hyperspectral image Perform anomaly detection on each spectral vector in to obtain the abnormal response value of each spectral vector;

[0128] The RX detector operator is calculated by formula (4):

[0129]

[0130] In formula (4), The fused hyperspectral image The spectral vector corresponding to the i-th row in , It is C B The inverse matrix of , the obtained RX(i) is the abnormal response value of the i-th spectral vector, and RX is the abnormal response matrix composed of the abnormal response values of all spectral vectors.

[0131] The above S150 is specifically:

[0132] Normalization is performed on the abnormal response matrix RX obtained in S140 to obtain an abnormality detection result map of the entire hyperspectral image. The implementation steps are as follows:

[0133] S151, performing normalization processing on the abnormal response matrix RX obtained in S140 to obtain a normalized abnormal response matrix Re, which is normalized using the following formula (5):

[0134]

[0135] In formula (5), min(RX) and max(RX) are the minimum and maximum values of all elements in the abnormal response matrix RX, respectively;

[0136] S152, the normalized abnormal response matrix Re is arranged according to the spatial position and converted into a two-dimensional abnormal response matrix Re of size H×W 2D , where H and W are the spatial height and spatial width of the original hyperspectral image, respectively;

[0137] S153, using Matlab to transform the two-dimensional abnormal response matrix Re 2D Displayed as an image, it is the anomaly detection result graph.

[0138] The embodiment of the present invention utilizes the GoDec algorithm to decompose a low-rank rough background matrix and a sparse rough anomaly matrix in the original hyperspectral image; utilizes the background network branch in the dual-branch network model to extract a pure, uniform, and Gaussian-distributed reconstructed background matrix from the input low-rank rough background matrix, and simultaneously utilizes the anomaly network branch in the dual-branch network model to extract a pure reconstructed anomaly matrix from the input sparse rough anomaly matrix; utilizes the separation enhancement constraint of the reconstruction space to explicitly enhance the spectral discrimination between the reconstructed background matrix and the reconstructed anomaly matrix; fuses the reconstructed background matrix and the reconstructed anomaly matrix to obtain fused hyperspectral data, and utilizes the reconstructed background matrix to calculate a pure background mean vector and a background covariance matrix; utilizes the calculated background mean vector and background covariance matrix, utilizes the RX detector operator to solve the anomaly response value of each spectral sample in the fused hyperspectral data, and thereby obtains an anomaly detection result map of the entire hyperspectral image.

[0139] Compared with existing hyperspectral anomaly detection methods, the unsupervised dual-branch network model in the embodiment of the present invention mines the deep spectral features of the background signal and the anomaly signal in parallel, and the design of the separability enhancement constraint term significantly improves the difference between the reconstructed background and the reconstructed anomaly, which is undoubtedly beneficial to anomaly detection. In addition, the embodiment of the present invention applies Gaussian adversarial training in the latent space of the background network branch, so that the background network branch outputs a pure and uniform background that satisfies the Gaussian distribution, solving the problem of mismatch between the background and the Gaussian distribution and the anomaly contamination problem in the RX algorithm, while the low-dimensional background manifold constraint in the latent space further utilizes the local structural features in the hyperspectral image, so that local correlation is preserved in the reconstructed data.

[0140] Example 2:

[0141] The experimental hyperspectral dataset used in this embodiment was collected by the Hyperspectral Digital Imagery Collection Experiment (HYDICE) sensor. It covers a suburban area measuring 80 x 100 x 160, with a spatial length and width of 80 x 100, respectively, and contains 160 spectral bands. This dataset is referred to as the Urban dataset.

[0142] Based on the solution disclosed in Example 1, please refer to Figure 2 , a hyperspectral anomaly detection method based on a dual-branch generative adversarial network in this embodiment includes:

[0143] Step 1: Convert the original hyperspectral dataset to be processed into a two-dimensional hyperspectral matrix and standardize it.

[0144] (a) The original hyperspectral dataset to be processed is converted into a two-dimensional hyperspectral matrix. The converted two-dimensional hyperspectral matrix X 2D The size of is N×S, where each row corresponds to each spectral band in the original hyperspectral dataset, and each column corresponds to the spectral vector of each pixel in the original hyperspectral dataset.

[0145] (b) The converted two-dimensional hyperspectral matrix is normalized by the following steps:

[0146] For a two-dimensional hyperspectral matrix X with size N×S 2D , calculate the mean μ of each column element i and variance δ i , and X 2D The following normalization formula is performed on each column element in:

[0147]

[0148] Where i=1,2,...,S represents the column index, X 2D (:,i) represents the two-dimensional hyperspectral matrix X 2D The i-th column vector in X is normalized. 2D The data of each band in is converted to a value with a mean of 0 and a variance of 1.

[0149] Step 2: Use GoDec to perform low-rank and sparse matrix decomposition to obtain a low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D .

[0150] The mathematical model of low-rank and sparse matrix decomposition is X 2D =B 2D +A 2D +N 2D , the specific decomposition steps of the GoDec algorithm are as follows:

[0151] (a) Set the rank upper bound r of the low-rank matrix, the sparsity upper bound k of the sparse matrix, and the error tolerance ε;

[0152] (b) Initialization: Let (0 represents a two-dimensional matrix of size N × S with all element values 0), where the superscript represents the number of iterations t = 0;

[0153] (c) If where ||·|| F Represents the F norm of the matrix, then stop the iteration and output and Execute step (j), otherwise, proceed to the next step;

[0154] (d) Let the number of iterations t = t + 1 and optimize the low-rank matrix B 2D , the formula is as follows:

[0155]

[0156] Where the power index q>0, is the optimized low-rank matrix;

[0157] (e) Constructing a low-rank matrix using a random matrix Bilateral random projection of :

[0158] Use the randn(S,r) and randn(N,r) commands to generate a random vector of size S×r and a random vector of size N×r, denoted as the right projection matrix M1 and the left projection matrix M2, and construct the right random projection and left random projection

[0159] (f) Update the left projection matrix M2 = Y1, and update the left random projection Update the right projection matrix M1 = Y2, and update the right random projection

[0160] (g) Perform QR decomposition on the bilateral random projections Y1 and Y2: Y1 = Q1R1 and Y2 = Q2R2;

[0161] (h) If rank(M2 T Y1) < r, then let r = rank(M2 T Y1), return to step (a), otherwise, continue to execute the next step;

[0162] (i) Update the matrix And update the matrix Where P Ω (·) represents the projection of the matrix onto the set Ω, and Ω is a non - zero subset composed of the first kN largest elements of , and return to step (c);

[0163] (j) Finally, And Are the low - rank rough background matrix B 2D And the sparse rough anomaly matrix A 2D .

[0164] Step 3: Use a double - branch network to reconstruct the low - rank rough background matrix and the sparse rough anomaly matrix in parallel:

[0165] (a) Input each spectral vector in the low - rank rough background matrix B 2D into the background network branch, and at the same time input each spectral vector in the sparse rough anomaly matrix A 2D into the anomaly network branch, and impose a separation - enhancement constraint on the reconstructed background matrix output by the background network branch and the reconstructed anomaly matrix output by the anomaly network branch. Iteratively train this double - branch network 2000 times to obtain a trained double - branch network model.

[0166] (a1) The background network branch consists of an encoder, a decoder, a Gaussian adversarial network in the latent layer, and a low - dimensional manifold constraint in the latent layer. The encoder in the background network branch receives each spectral vector in the low - rank rough background matrix B 2D , and learns the background distribution through the Gaussian adversarial network in the latent layer and the low - dimensional manifold constraint in the latent layer, extracts the intrinsic background spectral features, and finally the decoder outputs a pure and uniform reconstructed background matrix that satisfies the Gaussian distribution.

[0167] The structures corresponding to each module in the background network branch are as follows:

[0168] Encoder structure: input layer (the number of nodes is the number of bands S in the original hyperspectral image) → hidden layer (the number of nodes is 500) → hidden layer (the number of nodes is 500) → output layer (i.e., potential layer, the number of nodes is 20);

[0169] Decoder structure: input layer (i.e., potential layer, number of nodes is 20) → hidden layer (number of nodes is 500) → hidden layer (number of nodes is 500) → output layer (i.e., reconstruction layer, number of nodes is the number of bands S in the original hyperspectral image);

[0170] The Gaussian adversarial network structure of the latent layer is: input layer (i.e. latent layer, number of nodes is 20) → hidden layer (number of nodes is 500) → hidden layer (number of nodes is 500) → output layer (number of nodes is 1);

[0171] Low-dimensional manifold constraint of the potential layer: Using the local linear embedding LLE method, the low-rank rough background matrix B decomposed in step (2) is 2D Extract the low-dimensional background manifold B ML , which has a rough background matrix B with the encoder 2D The output potential data Z has the same dimension. Therefore, a low-dimensional manifold constraint is imposed on the potential layer, as follows:

[0172]

[0173] in is the low-dimensional background manifold B ML The spectral vector corresponding to the i-th row in z i is the i-th latent spectral vector in Z, and N is the number of pixels in the original hyperspectral image.

[0174] (a2) The abnormal network branch consists of an encoder and a decoder. The encoder in the abnormal network branch receives a sparse rough abnormal matrix A 2D For each spectral vector in , the decoder outputs a clean reconstructed anomaly matrix.

[0175] The structures corresponding to each module in the abnormal network branch are as follows:

[0176] Encoder structure: input layer (the number of nodes is the number of bands S in the original hyperspectral image) → hidden layer (the number of nodes is 500) → hidden layer (the number of nodes is 500) → output layer (i.e., potential layer, the number of nodes is 20);

[0177] Decoder structure: input layer (i.e., potential layer, the number of nodes is 20) → hidden layer (the number of nodes is 500) → hidden layer (the number of nodes is 500) → output layer (i.e., reconstruction layer, the number of nodes is the number of bands S of the original hyperspectral image);

[0178] (a3) The separation enhancement constraint is calculated by the following formula:

[0179]

[0180] in is the rough background matrix B of the decoder input in the background network branch 2D The reconstruction vector output by the i-th spectral vector in, is the rough anomaly matrix A of the decoder input in the anomaly network branch 2D The reconstructed vector output by the i-th spectral vector in .

[0181] Obviously, the above separability enhancement constraint in the reconstruction space can directly and explicitly enhance the spectral difference between the reconstructed background matrix and the reconstructed anomaly matrix.

[0182] (b) For the trained dual-branch network, extract the output of the decoder in the background network branch, which is the reconstructed background matrix Extract the output of the decoder in the abnormal network branch, which is the reconstructed abnormal matrix Step 4: Calculate the abnormal response value of the hyperspectral image.

[0183] Since the reconstructed background matrix obtained in step 3 above obeys a multivariate Gaussian distribution, and both the reconstructed background matrix and the reconstructed anomaly matrix are pure and have obvious spectral differences, it is very suitable to use a classic RX detector to perform anomaly detection on the hyperspectral dataset formed by the fusion of the reconstructed background matrix and the reconstructed anomaly matrix. Accordingly, this embodiment adopts the RX algorithm, and its specific processing steps are as follows:

[0184] (a) Reconstruct the background matrix and reconstruct the anomaly matrix Fusion is performed to obtain the fused hyperspectral image

[0185]

[0186] (b) Solve the reconstructed background matrix The mean vector and covariance matrix of all spectral vectors in , as the background mean vector μ B and the background covariance matrix C B , calculated using the following formula:

[0187]

[0188] in Represents the reconstructed background matrix The spectral vector corresponding to the i-th row in .

[0189] (c) According to the background mean vector μ B and the background covariance matrix C B , using the RX detector operator, the fused hyperspectral image Perform anomaly detection on each spectral vector in and obtain the abnormal response value of each spectral vector. It is calculated by the following formula:

[0190]

[0191] in The fused hyperspectral image The spectral vector corresponding to the i-th row in , It is C B The inverse matrix of , the obtained RX(i) is the abnormal response value of the i-th spectral vector, and RX is the abnormal response matrix composed of the abnormal response values of all spectral vectors.

[0192] Step 5: Normalize the abnormal response values of each spectral vector to obtain the abnormality detection result map of the entire hyperspectral image, which includes the following steps:

[0193] (a) Perform normalization processing on the abnormal response matrix RX obtained in step (4) to obtain the normalized abnormal response matrix Re. The normalization processing formula is as follows:

[0194]

[0195] Where min(RX) and max(RX) are the minimum and maximum values of all elements in the abnormal response matrix RX, respectively.

[0196] (b) The normalized abnormal response matrix Re is arranged according to the spatial position and converted into a two-dimensional abnormal response matrix Re of size H×W 2D , where H and W are the spatial height and spatial width of the original hyperspectral image, respectively;

[0197] (c) Using Matlab, the two-dimensional abnormal response matrix Re 2D Displayed as an image, it is the anomaly detection result graph.

[0198] Simulation experiment and result analysis:

[0199] Simulation experiment 1: Figure 3 This figure compares the anomaly detection results obtained by performing anomaly detection on an Urban hyperspectral dataset using the traditional RX anomaly detection method, the KRX anomaly detection method, the CRD anomaly detection method, and the anomaly detection method of an embodiment of the present invention.

[0200] in, Figure 3 (a) is a schematic diagram of the scene of the Urban hyperspectral dataset; Figure 3 (b) is the anomaly ground truth map of the scene, where black represents the background and white represents the anomaly; Figure 3 (c) is the anomaly detection result of the traditional RX anomaly detection method on the Urban dataset; Figure 3 (d) is the anomaly detection result of the KRX anomaly detection method on the Urban dataset; Figure 3 (e) is the anomaly detection result of the CRD anomaly detection method on the Urban dataset; Figure 3 (f) is the anomaly detection result diagram of the anomaly detection method of the present invention on the Urban dataset;

[0201] for Figure 3 The anomaly detection result diagrams shown in (c)-(f) show that brighter pixels correspond to stronger response values, while darker pixels represent weaker response values. It can be seen that compared with the traditional RX anomaly detection method, KRX anomaly detection method, and CRD anomaly detection method, the anomaly detection method of the present invention achieves more significant background-anomaly separation, in which the abnormal target is highlighted from the background with a higher degree of discrimination. The other three anomaly detection methods generally have a more serious false alarm problem, which makes it difficult to distinguish between the background and the anomaly.

[0202] Simulation experiment 2: Figure 4 This is a comparison chart of the ROC curves obtained by performing anomaly detection on the Urban hyperspectral dataset using the traditional RX anomaly detection method, the KRX anomaly detection method, the CRD anomaly detection method, and the anomaly detection method of an embodiment of the present invention.

[0203] It can be seen that compared with the traditional RX anomaly detection method, the KRX anomaly detection method and the CRD anomaly detection method, the ROC curve of the anomaly detection method in the embodiment of the present invention is closer to the upper left corner of the coordinate plane. Under the same false alarm rate, its detection rate is also significantly higher than that of other methods.

[0204] Simulation experiment 3: Figure 5 This is a comparison table of AUC values obtained by performing anomaly detection on the Urban hyperspectral dataset using the traditional RX anomaly detection method, the KRX anomaly detection method, the CRD anomaly detection method, and the anomaly detection method of an embodiment of the present invention.

[0205] Figure 5 It can be seen that compared with the traditional RX anomaly detection method, the KRX anomaly detection method and the CRD anomaly detection method, the anomaly detection method of the embodiment of the present invention has the highest AUC value, which proves that it has more accurate anomaly detection results.

[0206] In summary, compared with traditional hyperspectral anomaly detection methods, the anomaly detection method of the embodiment of the present invention has more superior anomaly detection performance. The embodiment of the present invention uses the background network branch and the anomaly network branch to extract spectral features and reconstruct spectral data for the low-rank rough background matrix and the sparse rough anomaly matrix, respectively, and imposes a separability enhancement constraint term in the reconstruction space to improve the spectral difference between the background and the anomaly. Through iterative training, the background network branch is able to output a pure, uniform reconstructed background matrix that satisfies the Gaussian distribution, while the anomaly network branch is able to output a pure reconstructed anomaly matrix. Finally, the background statistics are estimated from the reconstructed background matrix, and RX detection is performed on the fused hyperspectral dataset, effectively overcoming the mismatch problem between the background and the Gaussian distribution and the anomaly contamination problem in the background statistics faced by the traditional RX method. Therefore, the embodiment of the present invention can effectively improve the distinguishability between background and anomalies, and ultimately present very accurate anomaly detection results.

[0207] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

[0208] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0209] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A hyperspectral anomaly detection method based on a dual-branch generative adversarial network, characterized in that: include: Step 1: Decompose the low-rank rough background matrix B from the original hyperspectral image X 2D and the sparse rough anomaly matrix A 2D ; Step 2: The rough background matrix B 2D and the rough anomaly matrix A 2D Input the dual-branch network model, which corresponds to the rough background matrix B 2D Output reconstructed background matrix The two-branch network model corresponds to the rough anomaly matrix A 2D Output reconstructed anomaly matrix Step 3: Fusion of the reconstructed background matrix With the reconstructed anomaly matrix Obtain the fused hyperspectral data and reconstruct the background matrix using the Calculate the clean background mean vector and background covariance matrix; Step 4: using an RX detector operator to solve the abnormal response value of each spectral vector in the hyperspectral data according to the background mean vector and the background covariance matrix; Step 5: Normalize the abnormal response values of each spectral vector to obtain the abnormal detection result map of the entire hyperspectral image.

2. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 1 is characterized in that: The step 1 is specifically as follows: Step (11): convert the original hyperspectral image X into a two-dimensional hyperspectral matrix X with a size of N×S. 2D , and the two-dimensional hyperspectral matrix X 2D Perform standardization to obtain the standardized two-dimensional hyperspectral matrix X 2D ; Where: N is the number of pixels in X, S is the number of bands in X; Step (12): Using the GoDec algorithm, the normalized two-dimensional hyperspectral matrix X 2D Perform low-rank and sparse matrix decomposition to extract the low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D .

3. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 2 is characterized in that: The two-dimensional hyperspectral matrix X 2D Perform standardization processing, specifically: For a two-dimensional hyperspectral matrix X with size N×S 2D , calculate the mean μ of each column element corresponding to each band i and variance δ i , and X 2D The following normalization formula (1) is performed on each column element in: In formula (1), i=1,2,...,S represents the column index. After normalization, the two-dimensional hyperspectral matrix X 2D The data of each band in is converted to a value with a mean of 0 and a variance of 1.

4. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 2 is characterized in that: Using the GoDec algorithm, the normalized two-dimensional hyperspectral matrix X 2D Perform low-rank and sparse matrix decomposition to extract the low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D , specifically: (1) The mathematical model of low-rank and sparse matrix decomposition is X 2D =B 2D +A 2D +N 2D ; (2) The specific decomposition steps of the GoDec algorithm include: (2a) Set the rank upper bound r of the low-rank matrix, the sparsity upper bound k of the sparse matrix, and the error tolerance ε; (2b) Initialize the low-rank and sparse matrix: Where, the superscript represents the number of iterations t=0, and 0 represents a two-dimensional matrix of size N×S with all element values 0; (2c) If where ||·|| F Represents the F norm of the matrix, then stop the iteration and output and Execute step (2j), otherwise, proceed to the next step; (2d) Let the number of iterations t = t + 1, and use the following formula (2) to optimize the low-rank matrix B 2D : In formula (2), the power index q>0, is the optimized low-rank matrix; (2e) Constructing a low-rank matrix using a random matrix Bilateral random projection: Use the randn(S,r) and randn(N,r) commands to generate a random vector of size S×r and a random vector of size N×r, denoted as the right projection matrix M1 and the left projection matrix M2, and construct the right random projection and left random projection (2f) Update the left projection matrix M2 = Y1, update the left random projection Update the right projection matrix M1=Y2, update the right random projection (2g) Perform QR decomposition on the bilateral random projections Y1 and Y2: Y1 = Q1R1 and Y2 = Q2R2; (2h) If rank(M2 T Y1) < r, then set r = rank(M2 T Y1), return to step (2a), otherwise, continue to the next step; (2i) Update the matrix Update Matrix Among them, P Ω (·) represents the projection of the matrix onto the set Ω, which is composed of The non-zero subset consisting of the first kN largest elements of and return to step (2c); (2j)Finally, and That is the low-rank rough background matrix B 2D and the sparse rough anomaly matrix A 2D .

5. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 1 is characterized in that: The step 2 is specifically as follows: Step (21), the rough background matrix B 2D Each spectral vector in is input into the background network branch, and the rough anomaly matrix A is 2D Each spectral vector in is input into the abnormal network branch, and the reconstructed background matrix output by the background network branch is and the reconstructed anomaly matrix output by the anomaly network branch Apply the separation enhancement constraint and iteratively train the two-branch network model 2000 times to obtain the trained two-branch network model; Step (22): For the trained dual-branch network model, extract the output of the decoder in the background network branch to reconstruct the background matrix Extract the output of the decoder in the abnormal network branch, which is the reconstructed abnormal matrix 6. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 1 is characterized in that: Step three is as follows: Step (31), reconstructing the background matrix and the reconstructed anomaly matrix Fusion is performed to obtain the fused hyperspectral image Step (32), solve the reconstructed background matrix by formula (3) The mean vector and covariance matrix of all spectral vectors in are respectively denoted as background mean vector μ B and the background covariance matrix C B ; In formula (3), Represents the reconstructed background matrix The spectral vector corresponding to the i-th row in .

7. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 1 is characterized in that: The step 4 is specifically as follows: According to the background mean vector μ B and the background covariance matrix C B , using the RX detector operator, the fused hyperspectral image Perform anomaly detection on each spectral vector in to obtain the abnormal response value of each spectral vector; The RX detector operator is calculated by formula (4): In formula (4), The fused hyperspectral image The spectral vector corresponding to the i-th row in , It is C B The inverse matrix of , the obtained RX(i) is the abnormal response value of the i-th spectral vector, and RX is the abnormal response matrix composed of the abnormal response values of all spectral vectors.

8. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 1 is characterized in that: The step five is specifically as follows: Normalize the abnormal response matrix RX obtained in step 4 to obtain the abnormal detection result map of the entire hyperspectral image. The implementation steps are as follows: Step (51) performs normalization processing on the abnormal response matrix RX obtained in step 4 to obtain the normalized abnormal response matrix Re, which is normalized using the following formula (5): In formula (5), min(RX) and max(RX) are the minimum and maximum values of all elements in the abnormal response matrix RX, respectively; Step (52): Arrange the normalized abnormal response matrix Re according to the spatial position and convert it into a two-dimensional abnormal response matrix Re with a size of H×W. 2D , where H and W are the spatial height and spatial width of the original hyperspectral image, respectively; Step (53), use Matlab to transform the two-dimensional abnormal response matrix Re 2D Displayed as an image, it is the anomaly detection result graph.

9. The hyperspectral anomaly detection method based on a dual-branch generative adversarial network according to claim 5, characterized in that: In step (21): The background network branch consists of an encoder, a decoder, a Gaussian adversarial network of the potential layer, and a low-dimensional manifold constraint of the potential layer. The encoder in the background network branch receives a low-rank rough background matrix B 2D The spectral vectors in the latent layer are learned through the Gaussian adversarial network and the low-dimensional manifold constraint of the latent layer to extract the intrinsic background spectral features. Finally, the decoder outputs a pure and uniform reconstructed background matrix that satisfies the Gaussian distribution. The abnormal network branch consists of an encoder and a decoder. The encoder in the abnormal network branch receives a sparse rough abnormal matrix A 2D For each spectral vector in , the decoder outputs a clean reconstructed anomaly matrix; The separation enhancement constraint is calculated by the following formula: in is the rough background matrix B of the decoder input in the background network branch 2D The reconstruction vector output by the i-th spectral vector in, is the rough anomaly matrix A of the decoder input in the anomaly network branch 2D The reconstructed vector output by the i-th spectral vector in .

Images