Hyperspectral anomaly detection method based on spatial spectrum attention enhancement and blind spot U-Net network

Abstract

The invention discloses a hyperspectral anomaly detection method based on spatial spectrum attention enhancement and a blind spot U-Net network, the overall architecture of the network is in a U-Net form, and interference of an abnormal target on subsequent feature extraction is eliminated through mask convolution and down-sampling operation. Firstly, in order to effectively capture global information, a spectrum-space self-attention module is embedded in a network; particularly, for a spatial self-attention module, in order to prevent blind spot information from interfering subsequent feature extraction, a mask is added to an attention map to meet blind spot requirements. In addition, in order to perceive the multi-class attributes of the background, an attention module based on a sequence dictionary is designed in the potential layer space. The multi-class attributes of the background are perceived through an additional dictionary, so that the background expression is effectively enhanced, and the anomaly detection precision is improved. And after the network converges, obtaining an anomaly detection result through a reconstruction error. According to the method, target features can be enhanced for interference of small targets, complex backgrounds and the like, so that a relatively good detection effect is achieved.

Description

A Hyperspectral Anomaly Detection Method Based on Spatial Spectral Attention Enhancement and Blind Spot U-Net Network Technical Field

[0001] This invention belongs to the field of hyperspectral anomaly detection technology, specifically relating to a hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network. Background Technology

[0002] Hyperspectral imagery is a three-dimensional data cube, with two dimensions reflecting the spatial geometry of ground features and the other recording the reflectance intensity of objects at different wavelengths. Due to its rich spectral information, hyperspectral imagery can identify various ground features, providing data support for tasks such as precision agriculture, mineral exploration, and maritime search and rescue. Anomaly detection, in particular, aims to uncover potential targets by analyzing the distribution of given data without any prior information. Over the past few decades, researchers have developed numerous anomaly detection algorithms, which can be categorized based on their fundamental principles into statistical algorithms, representation-based algorithms, and deep learning-based detection methods.

[0003] Traditional methods mostly rely on mathematical statistics to design detectors using the spatial or spectral characteristics of images. They can be broadly categorized into statistical anomaly detection algorithms and representation-based anomaly detection algorithms. Statistical algorithms require assumptions about the background distribution, but in real-world scenarios, the background rarely conforms to a single distribution assumption, leading to low detection accuracy. Representation-based algorithms require setting a large number of parameters, which often vary significantly across different scenarios, making them difficult to apply in real-world tasks.

[0004] Unlike the methods mentioned above, deep learning-based methods do not require assumptions about the data distribution. Instead, they use networks to extract discriminative features or reconstruct the background for anomaly detection. Typically, anomalous pixels occupy only a small portion of an image, while background pixels make up the majority. Therefore, neural networks can encode and decode the input hyperspectral data, and then use the reconstruction error as the anomaly detection parameter.

[0005] Although the above methods have achieved good detection accuracy, they still have shortcomings in extracting global information. First, although some works use self-attention mechanisms to extract global information, blind spot information can easily be confused with subsequent feature extraction during feature interaction. In addition, the background usually has multi-class attributes, and effectively enhancing features of the same class is the key to further improving the quality of background reconstruction. Summary of the Invention

[0006] In view of this, the main objective of the present invention is to provide a hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network, so as to more effectively capture global information and ensure blind spot requirements, thereby improving the quality of background reconstruction and enhancing background features.

[0007] To achieve the above objectives, the technical solution of this invention is implemented as follows: a hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network, wherein the overall network architecture is U-Net, including the following steps: Step 1, using a layer of masked convolution to detect the input hyperspectral data Extracting features to obtain features , , , These represent the height, width, and number of spectral channels of the hyperspectral data, respectively, with N representing the number of channels of the feature; Step 2, ... The data is fed into the Spectral-Spatial Self-Attention (SSAB) module for global feature extraction to obtain the features. ,Will First, downsampling is performed, and then the samples are fed back into the SSAB module for feature extraction to obtain the features. The SSAB module mainly consists of a channel self-attention module based on dilated convolution, a feedforward network one, a multi-head spatial attention module based on masking, and a feedforward network two; Step 3, first... Downsampling is performed, and then the samples are fed into a sequence dictionary-based attention (TDAB) module for feature extraction to obtain features. The TDAB module mainly consists of a channel self-attention module based on dilated convolution, a third feedforward network, a cross-scale attention module based on sequence dictionaries, and a fourth feedforward network; Step 4, first... Perform upsampling, and then with Channel concatenation is performed, and then the concatenated features are dimensionality reduced. The dimensionality-reduced features are then fed into the SSAB module for feature extraction to obtain the final features. Step 5, first... Perform upsampling, and then with Channel concatenation is performed, and then the concatenated features are dimensionality reduced. The dimensionality-reduced features are then fed into the SSAB module for feature extraction to obtain the final features. Step 6, for The reconstructed background image is obtained by mapping. Step 7: After the network converges, the input image is fed into the network to restore the background, and anomaly detection is performed by using the reconstruction error.

[0008] In one embodiment, the SSAB module performs global feature extraction as follows: Input features are fed into a channel self-attention module based on dilated convolution, normalized, and then processed through dilated convolution and stretched to obtain the desired features. , , Matrix, pair Perform rotations and calculate the attention map for the channel direction, then... The matrices are multiplied, and then reshaped and skip connections are used to obtain output feature one. Output feature one is fed into feedforward network one to obtain output feature two. Output feature two is fed into a mask-based multi-head spatial attention module, normalized, and then dilated convolution is used to obtain... , , A matrix, where M is the window size, for The output feature three is obtained by transposing and calculating the spatial attention map. The masking strategy includes: only allowing even rows of the attention map to participate in the calculation, and only allowing diagonal elements to participate in the calculation. The output feature three is fed into the feedforward network two to obtain the output of the SSAB module.

[0009] In one embodiment, the channel self-attention module based on dilated convolution employs a 3×3 dilated convolution with a dilation rate of 2. The feedforward network one and feedforward network two first process the input features through an LN layer, then perform a 3×3 dilated convolution, finally perform non-linear processing using the GeLU activation function, and then again perform a 3×3 dilated convolution to obtain the output features. In one embodiment, step 3, the feature extraction steps of the TDAB module, are as follows: the input features are fed into the channel self-attention module based on dilated convolution, normalized, then processed by dilated convolution and stretched to obtain the output features. , , Matrix, pair Perform rotations and calculate the attention map for the channel direction, then... The matrices are multiplied, and then reshaped and skip connections are used to obtain output feature four. Output feature four is fed into feedforward network three to obtain output feature five. Output feature five is fed into a cross-scale attention module based on sequence dictionaries and transformed into query features through linear transformation. Obtain Key features from the dictionary and Value features Then, the features are enhanced through a self-attention mechanism, and the enhanced features are divided into n categories. Multi-head attention is then calculated on each category, and the results are finally merged to obtain output feature six. Output feature six is ​​fed into feedforward network four to obtain the output of the TDAB module.

[0010] In one embodiment, the sequence dictionary-based cross-scale attention module is divided into two sub-modules: a dictionary-based self-attention module (DSA) and an adaptive class multi-head attention module (CMA); the dictionary-based self-attention module is used to obtain... , and The size of the constructed dictionary is It is initialized using a Gaussian distribution, where P is the number of dictionary atoms; the adaptive category multi-head attention module is used to enhance features and obtain output feature six; spatial attention map. Represented as Enhanced features Represented as Based on spatial attention map A, the enhanced features will be... Divide into n categories, then calculate multi-head attention for each category, and finally merge the results to obtain output feature six.

[0011] Compared with the prior art, the beneficial effects of the present invention are: 1. The present invention proposes a U-Net hyperspectral anomaly detection network based on spectral-spatial self-attention enhancement. This network effectively captures global information through a spectral-spatial self-attention module and a sequence dictionary-based attention module, thereby accurately reconstructing the background.

[0012] 2. This invention introduces a spectral-spatial self-attention module, which can effectively mine global features of the background and ensure blind spot requirements by masking the attention map, thereby further improving the quality of background reconstruction.

[0013] 3. This invention introduces a sequence dictionary-based attention module in the latent space, which uses an additional dictionary to perceive multi-class background attributes and enhances background features by performing feature enhancement on the same category. Attached Figure Description

[0014] Figure 1 is a flowchart of the algorithm of the present invention.

[0015] Figure 2 is a schematic diagram of the spectral-spatial self-attention module, where a is the channel self-attention module based on dilated convolution and b is the multi-head spatial attention module based on mask.

[0016] Figure 3 is a schematic diagram of a cross-scale attention module based on sequence dictionaries, where a is a dictionary-based self-attention module (DSA) and b is an adaptive category multi-head attention module.

[0017] Figure 4 shows the detection results of each algorithm on the first dataset. From left to right and top to bottom, the results are: the labels corresponding to the dataset, the detection results of the RX algorithm, the detection results of the LREN algorithm, and the BS algorithm.3 The detection results of LNet algorithm, BockNet algorithm, PUNNet algorithm, NL2Net algorithm, DirectNet algorithm, BSDM algorithm, and the detection results of the algorithm of this invention.

[0018] Figure 5 shows the detection results of each algorithm on the second dataset. From left to right and top to bottom, the results are: the labels corresponding to the datasets, the detection results of the RX algorithm, the detection results of the LREN algorithm, and the BS algorithm. 3 The detection results of LNet algorithm, BockNet algorithm, PUNNet algorithm, NL2Net algorithm, DirectNet algorithm, BSDM algorithm, and the detection results of the algorithm of this invention.

[0019] Figure 6 shows the ROC curves of each algorithm on the second dataset. (a1) is the 3D ROC curve of each algorithm on the first dataset, and (a2) is the 2D ROC curve of each algorithm on the first dataset. (b1) is the 3D ROC curve of each algorithm on the second dataset, and (b2) is the 2D ROC curve of each algorithm on the second dataset.

[0020] Figure 7 shows the background-anomaly diagrams of each algorithm on the two datasets. (a) is the wireframe diagram of each algorithm on the first dataset, and (b) is the wireframe diagram of each algorithm on the second dataset. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] To further improve the global information extraction effect of hyperspectral images, avoid blind spot information from interfering with subsequent feature extraction, and effectively enhance features of the same category, this invention provides a hyperspectral anomaly detection method based on spatial spectral attention enhancement and a blind spot U-Net network. The overall network architecture is U-Net, and masked convolution and downsampling operations are used to eliminate the interference of anomalous targets on subsequent feature extraction.

[0023] First, to effectively capture global information, a spectral-spatial self-attention module is embedded in the network. Specifically, for the spatial self-attention module, to prevent blind spot information from interfering with subsequent feature extraction, a mask is added to the attention map to meet blind spot requirements. Furthermore, to perceive the multi-class attributes of the background, a sequence dictionary-based attention module is designed in the latent space. This module perceives the multi-class attributes of the background through an additional dictionary and further classifies the input features, learning their similarity for each class through a multi-head attention module. This effectively enhances background representation and further improves anomaly detection accuracy. Finally, after the network converges, the anomaly detection result is obtained by reconstructing the error. This invention can enhance target features and achieve better detection results, especially against interference from small targets and complex backgrounds.

[0024] Specifically, referring to Figure 1, the complete steps of the present invention are described as follows: Step 1, using a layer of masked convolution from the input hyperspectral data Extract features from them, , , These represent the height, width, and number of spectral channels of the hyperspectral data, respectively. The extracted features are represented as... N represents the number of channels of the feature, which is set to 48 in this invention.

[0025] Specifically, this step can be achieved by adding a batch normalization (BN) layer and a LeakyReLu activation function after the convolutional layer, as shown in the formula: ;in, For output features, This indicates a batch normalization operation. This represents the LeakyReLu activation function. Indicates 3 3. Masked convolution.

[0026] Step 2, The data is fed into the SSAB module for global feature extraction, and the extracted features are represented as follows: .

[0027] The SSAB module of this invention mainly consists of a channel self-attention module (D-CSA) based on dilated convolution, a feedforward network (FFN) I, a mask-based multi-head spatial attention module (M-MSA), and a feedforward network II.

[0028] For the D-CSA module, the input features are: First, normalization is performed using layer normalization (LN), then 3×3 dilated convolution with a dilation rate of 2 is applied followed by stretching to obtain the following results. , , Immediately afterwards, regarding Perform rotations and calculate the attention map for the channel direction, then... The matrices are multiplied together. Finally, the output feature is obtained by reshaping and skip connections, and then fed into the feedforward network (FFN).

[0029] The processing flow of the feedforward network (FFN) is as follows: the output feature one is first processed by an LN layer, and then fed into a 3×3 dilated convolution layer. The dilated convolution can effectively increase the receptive field, thereby extracting more local information. Finally, nonlinear processing is performed through the GeLU activation function, and the output feature two is obtained through a 3×3 dilated convolution layer, which is then fed into the M-MSA.

[0030] The above process can be represented as follows:

[0031] In the formula, and d represent the 3×3 dilated convolution with a dilation rate of 2 and the input feature dimension, respectively. and These represent output feature one obtained through D-CSA and output feature two obtained through FFN, respectively.

[0032] The M-MSA processing flow is as follows: the output feature 2 is first normalized by LN, and then 3×3 dilated convolution with a dilation rate of 2 is used to obtain the following results. , , A matrix, where M is the window size (set to 13 in this invention). Next, for... The output feature three is obtained by transposing the attention map and calculating the spatial attention map. The masking strategy mainly includes two steps. First, only even-numbered rows of the attention map are allowed to participate in the calculation; second, only diagonal elements are allowed to participate in the calculation, so the mask map is chessboard shaped. Output feature three is sent to FFN two. To prevent blind spot information from interfering with subsequent feature extraction, a mask is added to the attention map to meet the blind spot requirements. The masking strategy is as follows: only even-numbered rows of the attention map are allowed to participate in the calculation, and only diagonal elements are allowed to participate in the calculation; thus, the attention map after masking is chessboard shaped, which satisfies the blind spot calculation requirements.

[0033] The processing flow of FFN2 is the same as that of feedforward network 1, and the output features obtained after processing are the output results of the SSAB module.

[0034] Therefore, the mask-based multi-head attention mechanism is represented as: In the formula, For the mask matrix, This is the output of the SSAB module.

[0035] Step 3, First, downsampling is performed, and then the data is fed back into the SSAB module for feature extraction. The extracted features are represented as follows: .

[0036] In this step, downsampling is performed using PatchUnshuffle, as shown below: ;in, It is the result of step 3. and These represent the PatchUnshuffle process and the SSAB module, respectively.

[0037] The SSAB module used in this step has the same processing flow as the SSAB module in step 2. Considering processing efficiency, a separate SSAB module is used for both steps. In subsequent steps, a separate SSAB module will also be used.

[0038] Step 4, The feature representation generated is fed into the downsampling module. .

[0039] This downsampling step is implemented using PatchUnshuffle, as expressed in the formula: ; This indicates PatchUnshuffle processing.

[0040] Step 5, The feature representation generated is fed into the DTAB module. .

[0041] The sequence dictionary-based attention (TDAB) module of this invention mainly consists of D-CSA, FFN, a sequence dictionary-based cross-scale attention module (T-DCA), and FFN. The processing flow of D-CSA and FFN is the same as step 2, the difference being the initial input, which will not be described in detail here.

[0042] The T-DCA module is divided into two sub-modules: a dictionary-based self-attention module (DSA) and an adaptive class multi-head attention module (CMA). First, the input features... The feature representation after sequential processing through D-CSA and FFN is as follows: The dictionary built using a dictionary-based self-attention module (DSA) has a size of [size missing]. It is initialized using a Gaussian distribution, where P is the number of dictionary atoms, which is set to 256 in this invention. It transforms the input into query features through a linear transformation. Obtain Key features from the dictionary and Value features Specifically, to reduce computational cost, the dimensions of Query and Key are reduced to R. The above process can be represented as follows: ; In the formula, , , It is a linear transformation matrix.

[0043] Next, features are enhanced through a self-attention mechanism: In the formula, and Let A represent the attention map and the features after DSA enhancement, respectively. Based on the spatial attention map A, the input features can be divided into n categories. Then, multi-head attention is calculated on each category, and the results are finally merged to obtain the output features. CMA has two advantages: firstly, by calculating cross-attention within the same category, it can effectively reduce the computation of irrelevant features; secondly, it can effectively perceive global information, thus avoiding the receptive field limitations of window-based attention. Finally, the output features are obtained by passing through an FFN network. The above process can be represented as:

[0044] In the formula, These are the features output by the TDAB module. Categorize() and UnCategorize() represent classification and inverse classification processes, respectively.

[0045] Step 6, use the PatchShuffle operation to first... Perform upsampling, then combine with encoded features Channel concatenation is performed. Next, a 1×1 convolutional layer is used to reduce the dimensionality of the concatenated features. Finally, the dimensionality-reduced features are fed into an SSAB module for feature extraction, resulting in the following feature representation: .

[0046] The process of this step can be expressed by the following formula: In the formula, Concat[] represents the channel splicing operation. and These represent 1×1 convolution and upsampling, respectively. This indicates the SSAB module.

[0047] Step 7 is similar to step 6. Use the PatchShuffle operation to first... Perform upsampling, then combine with encoded features Channel concatenation is performed. Next, a 1×1 convolutional layer is used to reduce the dimensionality of the concatenated features. Finally, the dimensionality-reduced features are fed into an SSAB module for feature extraction, resulting in the following feature representation: .

[0048] The process of this step can be expressed by the following formula: Step 8: Use a single 1×1 convolutional layer. The reconstructed background image is obtained by mapping. .

[0049] The process of this step can be expressed by the following formula: In the formula, This is the reconstructed background image.

[0050] Step 9: After the network converges, the input image is fed back into the network to restore the background, and anomaly detection is performed by reconstructing the error.

[0051] The process of this step can be expressed by the following formula:

[0052] In the formula, This is the final 2D detection result image.

[0053] The effects of the present invention will be further described below with reference to simulation experiments.

[0054] 1. Simulation conditions: The hardware environment for the simulation experiment of this invention is an Intel Core i5-12400 CPU, 16-GB random access memory (RAM) and Microsoft Windows 10 operating system; the simulation software is MATLAB R2022a and Python 3.7.

[0055] 2. Experimental Content: To demonstrate the effectiveness of a hyperspectral anomaly detection method based on spatial spectral attention enhancement and a blind spot U-Net network, this invention uses two publicly available datasets to verify the algorithm's effectiveness and conducts simulation experiments. Figures 4 and 5 show example detection results of each algorithm on the two datasets. Figure 6 shows the ROC curves of each algorithm on the two datasets, and Figure 7 shows the background-anomaly diagrams of each algorithm.

[0056] In addition, Table 1 shows the AUC values ​​of different algorithms on the two datasets.

[0057] Table 1. AUC values ​​of different algorithms on two datasets. Algorithm | First Dataset | Second Dataset | RX | 0.9907 | 0.9946 | L | R | SR | 0.9539 | 0.9753 | BS 3LNet 0.9579 0.9470 BockNet 0.9891 0.9902 PUNNet 0.9833 0.9981 NL2Net 0.9893 0.9975 DirectNet 0.9881 0.9923 BSDM 0.9911 0.9971 Algorithm of this invention 0.9928 0.9954 Simulation experiments show that the present invention achieves good detection results in terms of both subjective vision and objective parameters. Specifically, the subjective detection results in Figures 4 and 5 demonstrate that the method of the present invention can accurately detect the position and shape of the target and effectively suppress the background. In contrast, the other comparison algorithms have varying degrees of missed detections and false alarms in the background, thus illustrating the superiority of the present invention. The ROC curves in Figure 6 show that the method of the present invention outperforms the other algorithms, indicating its superior overall performance. Furthermore, the background anomaly separation diagram in Figure 7 illustrates the relationship between background suppression and anomaly response. The present invention can effectively suppress the background while significantly detecting the target, demonstrating superior performance. Table 1 shows the AUC values ​​of each algorithm on the two datasets, indicating that the method of the present invention achieves the highest AUC value, demonstrating its effective target detection.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network, wherein the overall network architecture is U-Net, characterized in that, Includes the following steps: Step 1: Use a masked convolution layer to process the input hyperspectral data. Extracting features to obtain features ， 、 、 These represent the height, width, and number of spectral channels of the hyperspectral data, respectively, with N representing the number of channels of the feature; Step 2, ... The data is fed into a spectral-spatial self-attention module for feature extraction to obtain features. ,Will First, downsampling is performed, and then the sample is fed back into the spectral-spatial self-attention module for feature extraction to obtain the features. Step 3, first... Downsampling is performed, and then the samples are fed into a sequence dictionary-based attention module for feature extraction to obtain the features. Step 4, first... Perform upsampling, and then with Channel concatenation is performed, and then the concatenated features are dimensionality reduced. The dimensionality-reduced features are then fed into the SSAB module for feature extraction to obtain the final features. Step 5, first... Perform upsampling, and then with Channel concatenation is performed, and then the concatenated features are dimensionality reduced. The dimensionality-reduced features are then fed into the SSAB module for feature extraction to obtain the final features. Step 6, for The reconstructed background image is obtained by mapping. Step 7: After the network converges, the input image is fed into the network to restore the background, and anomaly detection is performed by using the reconstruction error.

2. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 1, characterized in that, Step 1 involves using a masked convolution layer to process the input hyperspectral data. Extracting features to obtain features , is represented as: ;in, This indicates a batch normalization operation. This represents the LeakyReLu activation function. Indicates 3 3. Masked convolution.

3. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 1, characterized in that, The spectral-spatial self-attention module includes a channel self-attention module based on dilated convolution, a feedforward network one, a mask-based multi-head spatial attention module, and a feedforward network two. Its feature extraction steps are as follows: input features are fed into the channel self-attention module based on dilated convolution, normalized, and then processed through dilated convolution and stretched to obtain the desired features. 、 ， Matrix, pair Transpose and calculate the attention map for the channel direction, then combine it with... The matrices are multiplied, and finally, the output feature is obtained by reshaping and skip connections. Output feature one is fed into feedforward network one to obtain output feature two; output feature two is fed into a mask-based multi-head spatial attention module, where it is normalized and then processed by dilated convolution to obtain... 、 、 A matrix, where M is the window size, for The output feature three is obtained by transposing and calculating the spatial attention map. The masking strategies include: only allowing even rows of the attention map to participate in the calculation, and only allowing diagonal elements to participate in the calculation. The output feature three is fed into the feedforward network two to obtain the output of the spectral-spatial self-attention module.

4. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 3, characterized in that, The channel self-attention module based on dilated convolution uses a 3×3 dilated convolution with a dilation rate of 2. The feedforward network one and feedforward network two first process the input features through an LN layer, then perform a 3×3 dilated convolution, finally perform nonlinear processing through the GeLU activation function, and then obtain the output features through a 3×3 dilated convolution again.

5. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 1, characterized in that, The sequence dictionary-based attention module includes a channel self-attention module based on dilated convolution, a third feedforward network, a cross-scale attention module based on the sequence dictionary, and a fourth feedforward network. Its feature extraction steps are as follows: input features are fed into the channel self-attention module based on dilated convolution, normalized, and then processed through dilated convolution and stretched to obtain the desired features. 、 ， Matrix, pair Perform rotations and calculate the attention map for the channel direction, then... The matrices are multiplied together, and finally, the output feature four is obtained through reshaping and skip connections. Output feature four is fed into feedforward network three to obtain output feature five; The output feature 5 is fed into a sequence dictionary-based cross-scale attention module, where it is transformed into query features through a linear transformation. Obtain Key features from the dictionary and Value features Then, the features are enhanced through a self-attention mechanism, the enhanced features are divided into n categories, multi-head attention is calculated on each category, and finally the results are merged to obtain the output feature six; The output feature six is ​​fed into the feedforward network four to obtain the output of the attention module based on the sequence dictionary.

6. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 5, characterized in that, The sequence dictionary-based cross-scale attention module is divided into two sub-modules: a dictionary-based self-attention module (DSA) and an adaptive class multi-head attention module (CMA). The dictionary-based self-attention module is used to obtain... ， and The size of the constructed dictionary is It is initialized with a Gaussian distribution, where P is the number of dictionary atoms; the adaptive category multi-head attention module is used to enhance features and obtain output feature six; Spatial attention map Represented as Enhanced features Represented as Based on spatial attention map A, the enhanced features will be... Divide into n categories, then calculate multi-head attention for each category, and finally merge the results to obtain output feature six.

7. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 6, characterized in that, The dictionary-based self-attention module reduces the dimensions of Query and Key to R, as follows: ； In the formula, ， ， It is a linear transformation matrix.

8. The method for detecting hyperspectral anomalies in a spatial-spectral attention enhancement and blind spot U-Net network according to claim 1, characterized in that, In step 2, PatchUnshuffle is used to... In step 3, downsampling is performed using PatchUnshuffle. Perform downsampling processing; in step 4, use PatchUnshuffle to... Upsampling is performed, and a 1×1 convolutional layer is used to reduce the dimensionality of the concatenated features; in step 5, PatchUnshuffle is used to... Upsampling is performed, and a 1×1 convolution layer is used to reduce the dimensionality of the concatenated features.

9. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 1, characterized in that, Step 6 uses a 1×1 convolutional layer. The reconstructed background image is obtained by mapping. 。 10. The hyperspectral anomaly detection method based on spatial spectral attention enhancement and blind spot U-Net network according to claim 1, characterized in that, Step 7, which involves anomaly detection based on reconstruction error, is represented as follows: In the formula, This is the final 2D detection result image.

Images