Hyperspectral image anomaly detection method based on end-to-end RX

Abstract

This invention discloses a hyperspectral image anomaly detection method based on end-to-end RX. The method comprises: using variational autoencoders to learn features for each pixel in the hyperspectral image; extracting hidden layer features of the variational autoencoder network as feature representations of the hyperspectral pixels; estimating the local mean and local covariance matrix of the feature representations of each pixel in the hyperspectral image; constructing a differentiable RX anomaly detection algorithm to obtain an anomaly index; calculating the backpropagation mode of the differentiable RX structure; constructing a hyperspectral anomaly detection network structure based on end-to-end RX; fusing the loss function of the variational autoencoder network and the end-to-end RX and training them; inputting a hyperspectral image into the trained network and outputting anomaly detection results. The method of the invention has the ability to effectively fuse the hidden layer feature space of the VAE network and the RX anomaly feature information, and has excellent performance when applied to hyperspectral image anomaly detection tasks.

Description

Technical Field

[0001] The present invention relates to a hyperspectral image anomaly detection technology, and in particular to a hyperspectral image anomaly detection method based on end-to-end RX. Background Art

[0002] Hyperspectral imagery (HSI) is a type of remote sensing data that contains both rich spectral and spatial information, typically represented as a three-dimensional cubic matrix. Given its ability to detect anomalous ground objects (small man-made objects such as aircraft and buildings) pixel by pixel, HSI has rapidly gained popularity in numerous fields, including geological surveys, environmental monitoring, urban planning, and crop assessment. Traditional anomaly detection algorithms for hyperspectral images include the collaborative representation detector (CRD), the attribute and edge-preserving detector (AED), and the low-rank and sparse representation detector (LRDM-MoG). With the rapid development of deep learning, HSI anomaly detection based on deep learning networks has become a key research topic in this field.

[0003] As a typical self-supervised deep network learning model, the autoencoder (AE) can perform dimensionality reduction and denoising on unstructured data such as images. It can extract the most representative information from the original data while reducing the amount of input data. Therefore, it can effectively extract spectral and spatial information from hyperspectral data and has attracted widespread attention from researchers. However, when using AE to model data, it is often difficult to distinguish the importance of different data. To improve the robustness and scalability of the model, researchers have proposed a variational autoencoder (VAE) for learning data distribution mapping, restricting the probability distribution of its latent vector to a standard normal distribution. As a generative model based on probabilistic statistics, the VAE can capture structural changes in data such as images and mine abstract features of high-dimensional data. Therefore, it can effectively extract spectral and spatial information from hyperspectral data. In hyperspectral anomaly detection algorithms based on AE and VAE, researchers typically use reconstruction error as an anomaly indicator for each pixel. Lu et al. [Lu X, Zhang W, Huang J. Exploiting Embedding Manifold of Autoencoders for Hyperspectral Anomaly Detection [J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(3): 1527-1537.] combined AE with manifold learning to design MC-AEN. By adding manifold constraints to the hidden layer, the intrinsic structure of the hyperspectral data was retained, and the global reconstruction error and the local implicit feature error were fused to obtain the final anomaly detection result. However, this type of method only emphasizes the approximation of the data and fails to fully utilize the abnormal target information in the hyperspectral data.

[0004] On the other hand, the traditional Reed-Xiaoli (RX) algorithm is a typical method for anomaly detection based on a probabilistic statistical model. It describes the background pixels in hyperspectral images using a Gaussian probability distribution model and calculates the Mahalanobis distance between each pixel and the background pixels, thereby distinguishing between anomaly pixels and background pixels. In the process of estimating statistical distribution parameters, some scholars have utilized the global information of hyperspectral images to construct a global RX (GRX) anomaly detection algorithm, which can effectively capture high-intensity anomaly targets in the entire region. Considering that there are still a large number of local anomalies or small target anomalies in hyperspectral images, local neighborhood information can also be used to estimate the statistical distribution parameters of background pixels. That is, a local RX (LRX) anomaly detection algorithm is adopted, which can better describe the background pixels in the local area and mine the smaller anomaly targets within this area. To improve the computational efficiency of the RX algorithm, Zheng et al. [Y Zheng, Y Li, Y Shi, et al. Acceleration scheme of RXD algorithm based on FPGA for hyperspectral anomaly target detection [J]. Beijing Hangkong Hangtian Daxue Xuebao / Journal of Beijing University of Aeronautics and Astronautics, 2018, 44(12): 2556-2567.] proposed a fast RX algorithm based on FPGA, which calculates the covariance matrix and its inverse matrix through block parallelization and QR decomposition. However, RX and its various derivative algorithms lack the ability to learn and cannot effectively capture the spectral and spatial information content of abnormal targets in different regions. Summary of the Invention

[0005] The present invention discloses a hyperspectral image anomaly detection method based on end-to-end RX. The method has the ability to extract abnormal pixels in hyperspectral images during the process of learning high-level features by variational autoencoder, and has excellent performance when applied to hyperspectral image anomaly detection tasks.

[0006] The technical solution to achieve the purpose of the present invention is: a method for detecting anomalies in a hyperspectral image based on end-to-end RX, comprising the following steps:

[0007] In the first step, a variational autoencoder network is used to learn the features of all pixels in the hyperspectral image to remove redundant information. That is, the high-dimensional hyperspectral data is used as the training data of VAE, and the stochastic gradient descent method is used to perform the VAE training process.

[0008] The second step is to extract the independent and identically distributed hidden layer data of each node as the feature representation of the hyperspectral pixel point based on the output data of the hyperspectral pixel point in the VAE hidden layer;

[0009] The third step is to estimate the local mean and local variance of the hidden layer feature representation data, that is, to perform mean filtering and square mean filtering on the hidden layer feature map of the hyperspectral image;

[0010] The fourth step is to construct a differentiable RX anomaly detection algorithm to obtain the anomaly index. This involves calculating the Mahalanobis distance on the hidden layer feature map of the VAE, performing K&I threshold segmentation after normalization to obtain the threshold parameter, and then calculating the anomaly index of each pixel.

[0011] The fifth step is to build a hyperspectral anomaly detection network structure based on end-to-end RX, that is, to build a network module that implements differentiable RX and embed the module into the VAE network;

[0012] The sixth step is to integrate the preliminary anomaly detection results with the loss function of the end-to-end RX, that is, to use the preliminary detection results as the weight coefficient of each pixel in the hyperspectral image in the end-to-end RX network;

[0013] In the seventh step, the network is trained using the stochastic gradient descent method and the final anomaly detection results are output.

[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the above-mentioned end-to-end RX-based hyperspectral image anomaly detection method is implemented.

[0015] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned end-to-end RX-based hyperspectral image anomaly detection method.

[0016] Compared with the existing technology, the present invention has the following significant features: (1) based on VAE, the mapping relationship of hyperspectral images from high-dimensional redundant data space to low-dimensional hidden layer independent distribution space is learned; (2) a differentiable RX anomaly detection module is established on the implicit feature map to obtain the anomaly index of each pixel; (3) the differentiable RX module is embedded in the VAE network to form an end-to-end RX-based anomaly detection network; (4) the network structure is simple, the computational complexity is small, and it has the advantages of fast training speed, high inference efficiency and high anomaly detection accuracy.

[0017] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a structural diagram of the method of the present invention.

[0019] Figure 2 These are the classification results of different methods on the HYDICE dataset, (a) pseudo color image. (b) Ground truth. (c) GRX. (d) LRX. (e) PCA-LRX. (f) CRD. (g) AED. (h) RX-VAE.

[0020] Figure 3 Classification results of different methods on the AVIRIS dataset. (a) Pseudo-color image. (b) Ground truth. (c) GRX. (d) LRX. (e) PCA-LRX. (f) CRD. (g) AED. (h) RX-VAE.

[0021] Figure 4 Classification results of different methods on the ABU-Urban Area dataset. (a) Pseudo-color image. (b) Ground truth. (c) GRX. (d) LRX. (e) PCA-LRX. (f) CRD. (g) AED. (h) RX-VAE.

[0022] Figure 5 Classification results of different methods on the ABU-Airport dataset. (a) Pseudo-color image. (b) Ground truth. (c) GRX. (d) LRX. (e) PCA-LRX. (f) CRD. (g) AED. (h) RX-VAE.

[0023] Figure 6 Classification results of different methods on the ABU-Beach dataset. (a) Pseudo-color image. (b) Ground truth. (c) GRX. (d) LRX. (e) PCA-LRX. (f) CRD. (g) AED. (h) RX-VAE.

[0024] Figure 7 It is the ROC curve on the ABU-airport dataset. DETAILED DESCRIPTION

[0025] In light of the challenges of existing anomaly detection algorithms, this paper proposes using VAE to learn the implicit features of independent and identically distributed (IID) hyperspectral data and designing a differentiable RX algorithm to filter anomaly information from the implicit feature graph. This method then guides the VAE to retain more anomaly information during training, thereby improving the anomaly detection performance of hyperspectral images. This method effectively integrates the advantages of VAE networks and traditional RX anomaly detection algorithms. While exploiting the intrinsic characteristics of high-dimensional and highly redundant hyperspectral data, it also retains anomaly information through an anomaly index weighting approach. The model exhibits high robustness, strong scalability, and excellent anomaly detection performance, making it widely applicable in related engineering fields. Experimental results demonstrate that the proposed method outperforms traditional methods in both detection results and detection accuracy on the benchmark datasets Hyperspectral Digital Imagery Collection Experiment (HYDICE), Airborne Visible / Infrared Imaging Spectrometer (AVIRIS), and Airport-Beach-Urban (ABU). The proposed method effectively integrates the hidden layer feature space of the VAE network and RX anomaly feature information, demonstrating excellent performance in anomaly detection tasks in hyperspectral images.

[0026] Combine Figure 1 The implementation process of the present invention is described in detail below. The steps are as follows:

[0027] The first step is to use the variational autoencoder network to learn the features of all pixels in the hyperspectral image and remove redundant information, that is, to use the high-dimensional hyperspectral data as the training data of the variational autoencoder network (VAE). is a pixel in the hyperspectral image, the matrix is the training sample set of all pixels, where B represents the number of bands of hyperspectral data and M×N represents the size of the hyperspectral image, i.e., height and width. First, the hyperspectral data is normalized to obtain Then, the hyperspectral pixel matrix The input is fed into the VAE network, where the encoder contains a 3×3 convolutional layer for capturing spatial features and two 1×1 convolutional layers for capturing spectral information, where the output of each node in the hidden layer is represented by [μ1, μ2, …, μ J ; σ1, σ2, …, σ J ] indicates that J corresponds to the dimension of the hidden layer feature representation. Finally, the output of the decoder is As the reconstructed data of the hyperspectral pixel point x, it is trained through the following loss function, specifically:

[0028]

[0029] The first and second terms correspond to the reconstruction error and KL divergence, respectively, σ j and μ j They represent the corresponding outputs of the hidden layer nodes respectively.

[0030] The second step is to extract the independent and identically distributed hidden layer data of each node as the feature representation of the hyperspectral pixel point based on the output data of the hyperspectral pixel point in the VAE hidden layer. First, the encoder and decoder of the VAE are respectively φ (z|x) and p θ (x|z) represents the feature representation of the hidden layer, and the subscripts φ and Represent the network parameters of the encoder and decoder respectively. The high-level feature representation z of the hyperspectral pixel is obtained by the output of the hidden layer node, specifically

[0031] z=μ+σ⊙ε'

[0032] Where μ=[μ1,μ2,…,μ J ],σ=[σ1,σ2,…,σ J ] is the output of the hidden layer node, ε~N(0,I) obeys the J-dimensional standard normal distribution.

[0033] The third step is to estimate the local mean and local variance of the hidden layer feature representation data, that is, to perform mean filtering and square mean filtering on the hidden layer feature map of the hyperspectral image. First, the high-level feature map representation of the hyperspectral pixel is denoted as Z 2 Represents the element-wise square operation of the matrix. Then, Z and Z are 2 Perform a mean filter operation with a window size of 7×7. Specifically, for each element z in any z∈Z l ∈z∈Z, the corresponding representations of its neighborhood pixels are z l (k), k=1, 2, ..., K, the local mean is represented by E[z l ] represents the local covariance, The local variance is expressed as var( z l) indicates that the specific calculation method is

[0034]

[0035]

[0036]

[0037] Wherein, K represents the number of neighborhood pixels, where K=7×7=49, and the local variance is obtained by subtracting the square of the local mean from the square of the local mean.

[0038] The fourth step is to construct a differentiable RX anomaly detection algorithm to obtain the anomaly index, that is, to calculate the Mahalanobis distance on the hidden layer feature map, perform K&I threshold segmentation after normalization to obtain the threshold parameter, and then calculate the anomaly index of each pixel. Apply the RX detection algorithm on the high-level feature map Z of the hyperspectral data to obtain a differentiable RX detection algorithm, specifically:

[0039]

[0040] Among them, d RX represents the Mahalanobis distance, The local mean of the corresponding background pixels, C b The covariance matrix corresponding to the background pixels, represents the eigenvector with zero mean.

[0041] The fifth step is to build a hyperspectral anomaly detection network structure based on end-to-end RX, that is, to build a network module that implements differentiable RX and embed the module into the VAE network. First, let Mean(·;7) be the mean filter function with a window size of 7×7. Then, the differentiable RX detection algorithm is implemented using the filter function of local mean and squared local mean, specifically

[0042]

[0043] Next, the mean filter function is embedded into the VAE network according to this formula, the output layer of the RX detection algorithm is normalized and K&I threshold segmentation is performed to obtain the threshold parameter θ, then the abnormal index of each pixel is related to d RX is proportional to

[0044]

[0045] Among them, h a The anomaly index of each pixel is represented by the value of RX. The larger the value, the higher the probability of it being an anomaly. Finally, the normalization layer and the anomaly index layer are connected to the VAE network to form an end-to-end RX anomaly detection network.

[0046] The sixth step is to integrate the initial anomaly detection results and the loss function of the end-to-end RX, that is, to use the initial detection results as the weight coefficients of each pixel in the hyperspectral image in the end-to-end RX network. First, the anomaly index h of each pixel is a The initial detection results are used as weight coefficients for each pixel in the VAE network training process. Then, an abnormality exponential weighted loss function is constructed to guide the VAE to retain more abnormal information.

[0047] Specifically

[0048]

[0049] Among them, the anomaly index of each hyperspectral pixel x is the corresponding h a , w and W′ are the network parameters of the encoder and decoder in VAE respectively.

[0050] The seventh step is to use the stochastic gradient descent method to train the network and output the final anomaly detection results. First, the encoder parameters and decoder parameters of the trained network are W and 终 and W 终 ′, then the encoder function is f(·; W 终 ), the output of the encoder is μ 终 =[μ1, μ2, …, μ J ] and σ final=[σ1,σ2,…,σ J ] cascade, that is

[0051] [μ 终 ,σ 终 ]=[μ1,μ2,…,μ J ,σ1,σ2,…,σ J ]=f(·;W 终 )

[0052] Then, the high-level features of the hyperspectral data in the hidden layer are expressed by the formula z 终 =μ 终 +σ 终 ⊙ε is sampled and input into the differentiable RX layer to finally obtain the abnormal index h of each pixel in the hyperspectral image a , forming the final anomaly detection result graph H a =[h a ] M×N .

[0053] The effects of the present invention can be further illustrated by the following simulation experiments: Hyperspectral Digital Imagery Collection Experiment (HYDICE), Airborne Visible / Infrared Imaging Spectrometer (AVIRIS) and Airport-Beach-Urban (ABU).

[0054] Simulation conditions

[0055] The simulation experiment uses three sets of real hyperspectral data: the HYDICE dataset, the AVIRIS dataset, and the ABU dataset.

[0056] The hyperspectral images in the HYDICE dataset contain 175 bands, with a spectral resolution of 10 nm and a spatial resolution of 1 m. The images are sized 80 × 100 pixels and cover areas such as vegetable fields, built-up areas, roads, and cars. The AVIRIS dataset, acquired by the airborne visible infrared imaging spectrometer at the San Diego Experimental Area in the United States, includes two 100 × 100 hyperspectral images with 189 bands. The ABU dataset contains 13 hyperspectral images with 100 × 100 or 150 × 150 pixels and 191, 205, 102, 188, 191, 193, 204, 205, and 207 bands, respectively. The images cover areas such as airports, beaches, and urban areas.

[0057] For the HYDICE, AVIRIS, and ABU datasets, the experiments mainly present anomaly detection result graphs, and list the ROC curves of some detection result graphs. The ROC curve describes the false positive rate (FPR, P d ), true positive rate (TPR, P f ) and the threshold (τ). In the experiment, the method of the present invention is recorded as RX-VAE, and the comparison methods are GRX, LRX, PCA-LRX, CRD and AED. The simulation experiments are all completed under the Windows 10 operating system using Python-3.6+Tensorflow-1.12+Pytorch-1.1. In addition, in the experiment of the present invention, the encoder and decoder of the VAE network each contain two hidden layers, the number of nodes in the decoder input layer and the number of nodes in the encoder output layer correspond to the number of spectral segments of each hyperspectral image, the activation function of each hidden layer uniformly adopts Leaky ReLU, and the network is trained by stochastic gradient descent. The learning rate is set to 0.0005, 0.0002 and 0.0005 in the three data sets, and the number of iterations is set to 200, 500 and 600, respectively.

[0058] Analysis of simulation experiment results

[0059] Figure 2 and Figure 3 The anomaly detection results of the simulation experiments on HYDICE and AVIRIS datasets by the method of the present invention and its comparative method are shown respectively. Figure 4-Figure 6 The following are the anomaly detection result diagrams of three different scenes in the ABU dataset: urban area, airport, and beach. From the experimental results, the present invention has achieved good results on the three different datasets. From the anomaly detection result diagrams of different algorithms, it can be seen that the detection result diagram of the method of the present invention can more clearly display the abnormal target and suppress the background more cleanly. A hyperspectral image is selected from each of the three datasets to show the ROC curves of the detection results of different comparison methods and the method of the present invention, as shown in the figure. Figure 7As shown. Figure 7 The ROC curves show that the quantitative indicators of the detection results of the method of the present invention are better than those of the comparison algorithms. The simulation experimental results of the above three different sets of real hyperspectral image data sets on the comparison algorithms and the method of the present invention demonstrate the effectiveness of the method of the present invention.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A hyperspectral image anomaly detection method based on end-to-end RX, characterized in that: The following steps are involved: In the first step, a variational autoencoder network is used to learn the features of all pixels in the hyperspectral image to remove redundant information. That is, the high-dimensional hyperspectral data is used as the training data of VAE, and the stochastic gradient descent method is used to perform the VAE training process. The second step is to extract the independent and identically distributed hidden layer data of each node as the feature representation of the hyperspectral pixel point based on the output data of the hyperspectral pixel point in the VAE hidden layer; The third step is to estimate the local mean and local variance of the hidden layer feature representation data, that is, to perform mean filtering and square mean filtering on the hidden layer feature map of the hyperspectral image; The fourth step is to construct a differentiable RX anomaly detection algorithm to obtain the anomaly index. This involves calculating the Mahalanobis distance on the hidden layer feature map of the VAE, performing K&I threshold segmentation after normalization to obtain the threshold parameter, and then calculating the anomaly index of each pixel. The fifth step is to build a hyperspectral anomaly detection network structure based on end-to-end RX, that is, to build a network module that implements differentiable RX and embed the module into the VAE network, as follows: Let Mean(·;7) be the mean filter function with a window size of 7×7; then, the differentiable RX detection algorithm is implemented using the filter functions of local mean and squared local mean, specifically: The high-level feature map representation of the hyperspectral pixel is denoted as M×N represents the size of the hyperspectral image, J corresponds to the dimension of the hidden layer feature representation, z l ∈z∈Z, z is the feature representation of the hidden layer; The mean filter function is embedded into the VAE network according to this formula, the output layer of the RX detection algorithm is normalized and K&I threshold segmentation is performed to obtain the threshold parameter θ, then the abnormal index of each pixel is related to d RX is proportional to Among them, h a Indicates the abnormality index of each pixel; Connect the normalization layer and the anomaly index layer to the VAE network to form an end-to-end RX anomaly detection network; The sixth step is to integrate the preliminary anomaly detection results with the loss function of the end-to-end RX, that is, to use the preliminary detection results as the weight coefficient of each pixel in the hyperspectral image in the end-to-end RX network; In the seventh step, the network is trained using the stochastic gradient descent method and the final anomaly detection results are output.

2. The end-to-end RX-based hyperspectral image anomaly detection method according to claim 1, characterized in that: The first step is to use the variational autoencoder network to learn the features of all pixels in the hyperspectral image and remove redundant information, that is, to use the high-dimensional hyperspectral data as the training data of VAE; record the column vector is a pixel in the hyperspectral image, the matrix is the training sample set of all pixels, where B represents the number of bands of hyperspectral data, and M×N represents the size of the hyperspectral image, i.e., height and width. First, the hyperspectral data is normalized to obtain Then, the hyperspectral pixel matrix The input is fed into the VAE network, where the encoder contains a 3×3 convolutional layer for capturing spatial features and two 1×1 convolutional layers for capturing spectral information. The output of each node in the hidden layer is represented by [μ1,μ2,…,μ J ;σ1,σ2,…,σ J ] indicates that J corresponds to the dimension of the hidden layer feature representation; finally, the output of the decoder As the reconstructed data of the hyperspectral pixel point x, it is trained through the following loss function, specifically: The first and second terms correspond to the reconstruction error and KL divergence, respectively, σ j and μ j They represent the corresponding outputs of the hidden layer nodes respectively.

3. The end-to-end RX-based hyperspectral image anomaly detection method according to claim 2, characterized in that: In the second step, based on the output data of the hyperspectral pixel in the VAE hidden layer, the independent and identically distributed hidden layer data of each node is extracted as the feature representation of the hyperspectral pixel; the details are as follows: The encoder and decoder of VAE are respectively q φ (z|x) and p θ (x|z) represents the feature representation of the hidden layer, and the subscripts φ and Represent the network parameters of the encoder and decoder respectively; the high-level feature representation z of the hyperspectral pixel is obtained by the output of the hidden layer node, specifically z=μ+σ⊙ε, where μ=[μ1,μ2,…,μ J ],σ=[σ1,σ2,…,σ J ] is the output of the hidden layer node, Obey the J-dimensional standard normal distribution.

4. The end-to-end RX-based hyperspectral image anomaly detection method according to claim 3, characterized in that: The third step is to estimate the local mean and local variance of the hidden layer feature representation data, that is, to perform mean filtering and square mean filtering on the hidden layer feature map of the hyperspectral image; the details are as follows: The high-level feature map representation of the hyperspectral pixel is denoted as Z 2 Represents the element-wise square operation of a matrix; For Z and Z respectively 2 Perform a mean filter operation with a window size of 7×7; for each element z in any z∈Z l ∈z∈Z, the corresponding representations of its neighborhood pixels are z l (k), k=1,2,...,K, local mean is E[z l ] represents the local covariance, The local variance is expressed as var(z l ) indicates that the specific calculation method is Wherein, K represents the number of neighborhood pixels, where K=7×7=49, and the local variance is obtained by subtracting the square of the local mean from the square of the local mean.

5. The end-to-end RX-based hyperspectral image anomaly detection method according to claim 4, characterized in that: The fourth step is to construct a differentiable RX anomaly detection algorithm to obtain the anomaly index. That is, the Mahalanobis distance is calculated on the hidden layer feature map, and after normalization, the K&I threshold segmentation is performed to obtain the threshold parameter, and then the anomaly index of each pixel is calculated. The RX detection algorithm is applied to the high-level feature map Z of the hyperspectral data to obtain a differentiable RX detection algorithm. Specifically, Among them, d RX represents the Mahalanobis distance, The local mean of the corresponding background pixels, C b The covariance matrix corresponding to the background pixels, represents the eigenvector with zero mean.

6. The end-to-end RX-based hyperspectral image anomaly detection method according to claim 5, characterized in that: The sixth step is to integrate the preliminary anomaly detection results with the end-to-end RX loss function, that is, to use the preliminary detection results as the weight coefficients of each pixel in the hyperspectral image in the end-to-end RX network; the details are as follows: The abnormality index h of each pixel a As the preliminary detection result, it is used as the weight coefficient of each pixel in the VAE network training process; Construct an abnormal exponential weighted loss function to guide VAE to retain more abnormal information; specifically Among them, the anomaly index of each hyperspectral pixel x is the corresponding h a , W and W′ are the network parameters of the encoder and decoder in VAE respectively.

7. The end-to-end RX-based hyperspectral image anomaly detection method according to claim 6, characterized in that: Step 7: Use stochastic gradient descent to train the network and output the final anomaly detection results; the details are as follows: The encoder parameters and decoder parameters in the trained network are W 终 and W 终 ′, then the encoder function is f(·; W 终 ), the output of the encoder is μ 终 =[μ1,μ2,…,μ J ] and σ 终 =[σ1,σ2,…,σ J ] cascade, that is [m 终 ,s 终 ]=[μ1,μ2,…,μ J ,σ1,σ2,…,σ J ]=f(·;W 终 ) The high-level features of hyperspectral data in the hidden layer are expressed by formula z 终 =μ 终 +σ 终 ⊙ε is sampled and input into the differentiable RX layer to finally obtain the abnormal index h of each pixel in the hyperspectral image a , forming the final anomaly detection result graph H a =[h a ] M×N .

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the end-to-end RX-based hyperspectral image anomaly detection method according to any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the end-to-end RX-based hyperspectral image anomaly detection method according to any one of claims 1 to 7 is implemented.

Images