Hyperspectral image anomaly detection method and system based on Mama network model

Abstract

The invention discloses a hyperspectral image anomaly detection method and system based on a Mama network model, relates to the technical field of hyperspectral remote sensing image processing, and is applied to a Mama network model comprising a preset spectral cluster unfolding assembly, a preset non-local reconstruction feature extraction assembly and a preset spectral cluster folding assembly. Comprising the following steps: processing an original hyperspectral image by using a preset spectral cluster expansion component to obtain a to-be-reconstructed spectral cluster set; performing non-local feature extraction and background reconstruction feature learning on the to-be-reconstructed spectral cluster set by using a preset non-local reconstruction feature extraction component to obtain a first reconstructed spectral cluster set; processing the first reconstructed spectral cluster set by using a preset spectral cluster folding component to obtain a second reconstructed spectral cluster set; an anomaly detection result map is determined based on the second set of reconstructed spectral clusters. Detection performance degradation caused by non-local feature extraction limitation and abnormal target pollution can be prevented.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing image processing, and in particular to a hyperspectral image anomaly detection method and system based on a Mamba network model. Background Art

[0002] The target detection technology of hyperspectral images can filter out candidate target pixels by modeling the spectral difference between the target and the background, and perform threshold segmentation on the candidate target pixels to obtain the area of ​​suspected targets. Among them, when the target characteristics are not fully understood, the spectral information of the target is in an unknown state. At this time, it is necessary to use hyperspectral anomaly detection technology to achieve "blind" detection of small targets, and then manually screen and label the target area, and establish a target sample library based on the labeled target area to provide key information for subsequent target characteristic analysis and recognition judgment.

[0003] At present, there are many high-performance hyperspectral anomaly detection algorithms on the market, such as anomaly detection methods based on data statistical distribution, collaborative and sparse representation models, low-rank sparse decomposition, and deep neural network models. However, model-based methods need to optimize model parameters according to the background and target characteristics of hyperspectral images, and rely too much on manual features, which limits the performance of anomaly detection. In addition, the hyperspectral image anomaly detection method based on the Mamba network model based on the deep neural network model realizes anomaly detection by deep representation learning and difference feature extraction of spectral sample data, but most of these methods have the problems of limited non-local feature extraction and high computational complexity, which limits the network detection performance of anomaly detection in the reconstruction learning mode.

[0004] From the above, it can be seen that how to prevent the detection performance degradation accidents caused by limited non-local feature extraction and abnormal target contamination in the background reconstruction learning mode in the process of hyperspectral image anomaly detection based on the Mamba network model is an urgent problem to be solved. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a hyperspectral image anomaly detection method and system based on the Mamba network model, which can prevent the problem of detection performance degradation caused by limited non-local feature extraction and abnormal target contamination in the background reconstruction learning mode during the hyperspectral image anomaly detection based on the Mamba network model. The specific scheme is as follows:

[0006] In a first aspect, the present application provides a hyperspectral image anomaly detection method based on a Mamba network model, which is applied to a Mamba network model including a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component, and a preset spectral cluster folding component, and the method includes:

[0007] The original hyperspectral image is processed in sequence by using a preset sliding window spectral cluster expansion component and a first local space and spectral feature compensation component in the preset spectral cluster expansion component to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image;

[0008] Using the preset non-local reconstruction feature extraction component, respectively perform non-local feature extraction and background reconstruction feature learning on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set, so as to obtain a first reconstructed spectrum cluster set;

[0009] Using the second local space and spectral feature compensation component in the preset spectral cluster folding component and the preset sliding window spectral cluster folding component to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain a second reconstructed spectral cluster set;

[0010] An anomaly detection result map corresponding to the original hyperspectral image is determined based on the second reconstructed spectrum cluster set.

[0011] Optionally, the using of a preset sliding window spectral cluster expansion component and a first local space and spectral feature compensation component in the preset spectral cluster expansion component to sequentially process the original hyperspectral image to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image includes:

[0012] Utilizing a preset sliding window spectral cluster expansion component in the preset spectral cluster expansion component to perform spectral cluster expansion on the original hyperspectral image, so as to obtain a plurality of spectral clusters to be processed corresponding to the original hyperspectral image; the sizes of the spectral clusters to be processed are the same;

[0013] In the first local space and spectral feature compensation component, a plurality of depth-separable convolution branches are used to extract features of each of the spectral clusters to be processed, so as to obtain corresponding first local space and spectral neighborhood features; the convolution kernel sizes corresponding to the depth-separable convolution branches are different;

[0014] Using preset splicing and convolution rules, each of the first local space and spectral neighborhood features is spliced ​​and convolved to obtain first splicing features corresponding to each of the first local space and spectral neighborhood features;

[0015] In the first local space and spectral feature compensation component, a plurality of depth-separable convolution branches are again used to extract corresponding second local space and spectral neighborhood features from each of the first spliced ​​features, and each of the second local space and spectral neighborhood features is spliced ​​and convolved using a preset splicing and convolution rule to obtain second splicing features corresponding to each of the second local space and spectral neighborhood features;

[0016] Each of the to-be-processed spectrum clusters is fused with the corresponding second splicing feature by using a preset residual connection to obtain a plurality of to-be-reconstructed spectrum clusters and a to-be-reconstructed spectrum cluster set including each of the to-be-reconstructed spectrum clusters.

[0017] Optionally, the using the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set to obtain a first reconstructed spectrum cluster set includes:

[0018] In the preset non-local reconstruction feature extraction component, a conversion operation is performed on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set to obtain a corresponding initial two-dimensional sequence feature;

[0019] Using the first convolution layer, the root mean square layer normalization layer and the linear transformation layer in the preset non-local reconstruction feature extraction component to process each of the initial two-dimensional sequence features in turn, to obtain the corresponding two-dimensional sequence features to be processed;

[0020] Inputting each of the two-dimensional sequence features to be processed into the first branch, so as to process each of the two-dimensional sequence features to be processed respectively by using the second convolutional layer, the first preset activation function and the preset visual state space model in the first branch to obtain a corresponding first processing result;

[0021] Inputting each of the two-dimensional sequence features to be processed into the second branch, so as to process each of the two-dimensional sequence features to be processed respectively using a second preset activation function in the second branch to obtain a corresponding second processing result;

[0022] The first processing result, the second processing result and the initial two-dimensional sequence feature are fused by using the bottleneck residual connection in the preset non-local reconstruction feature extraction component to obtain a corresponding fusion result, and each of the fusion results is respectively subjected to an inverse transformation operation to obtain a first reconstructed spectral cluster set.

[0023] Optionally, the step of inputting each of the two-dimensional sequence features to be processed into the first branch so as to use the second convolutional layer, the first preset activation function and the preset visual state space model in the first branch to process each of the two-dimensional sequence features to be processed respectively to obtain a corresponding first processing result includes:

[0024] Inputting each of the two-dimensional sequence features to be processed into the first branch, so as to use the second convolution layer and the first preset activation function in the first branch to process each of the two-dimensional sequence features to be processed respectively, so as to obtain corresponding features to be fused;

[0025] In the preset visual state space model, each of the features to be fused is scanned using a preset four-way scanning strategy, and the scanned results are fused to obtain a corresponding first processing result; the first processing result is a non-local feature corresponding to each of the spectral clusters to be reconstructed.

[0026] Optionally, the method of using the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain a second reconstructed spectral cluster set includes:

[0027] Using the third convolution layer in the preset spectrum cluster folding component and the second local space and spectrum feature compensation component to sequentially process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set, so as to obtain a corresponding compensated spectrum cluster;

[0028] A preset sliding window spectrum cluster folding component is used to perform a folding operation on each of the compensated spectrum clusters to obtain a second reconstructed spectrum cluster set.

[0029] Optionally, after the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component are used to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain the second reconstructed spectral cluster set, the method further includes:

[0030] A preset reconstruction loss calculation function is used to perform reconstruction loss calculation on the second reconstructed spectrum cluster set and the to-be-reconstructed spectrum cluster set to obtain corresponding reconstruction loss results, so as to use the reconstruction loss results to adjust the model parameters of the Mamba network model.

[0031] Optionally, determining an anomaly detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectrum cluster set includes:

[0032] Determine, by using a preset reconstruction residual technique and based on the second reconstructed spectrum cluster set, a reconstruction residual graph corresponding to each of the reconstructed spectrum clusters in the second reconstructed spectrum cluster set;

[0033] The reconstructed residual images are fused using a preset fusion technology to obtain an anomaly detection result image corresponding to the original hyperspectral image.

[0034] In a second aspect, the present application provides a hyperspectral image anomaly detection system based on a Mamba network model, which is applied to a Mamba network model including a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component, and a preset spectral cluster folding component, and the system includes:

[0035] A first spectrum cluster set determination module is used to sequentially process the original hyperspectral image using a preset sliding window spectrum cluster expansion component and a first local space and spectrum feature compensation component in the preset spectrum cluster expansion component to obtain a spectrum cluster set to be reconstructed corresponding to the original hyperspectral image;

[0036] A second spectrum cluster set determination module is used to use the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the spectrum clusters to be reconstructed in the spectrum cluster set to be reconstructed, so as to obtain a first reconstructed spectrum cluster set;

[0037] a third spectrum cluster set determination module, configured to sequentially process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set by using the second local space and spectrum feature compensation component and the preset sliding window spectrum cluster folding component in the preset spectrum cluster folding component to obtain a second reconstructed spectrum cluster set;

[0038] The detection result graph determining module is used to determine an abnormality detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectrum cluster set.

[0039] Optionally, the first spectrum cluster set determination module includes:

[0040] a spectral cluster expansion unit, configured to perform spectral cluster expansion on the original hyperspectral image using a preset sliding window spectral cluster expansion component in the preset spectral cluster expansion component, so as to obtain a plurality of spectral clusters to be processed corresponding to the original hyperspectral image; the sizes of the spectral clusters to be processed are the same;

[0041] A feature extraction unit, used for performing feature extraction on each of the spectral clusters to be processed by using a plurality of depth-separable convolution branches in the first local space and spectral feature compensation component to obtain corresponding first local space and spectral neighborhood features; the convolution kernel sizes corresponding to the depth-separable convolution branches are different;

[0042] A first splicing feature acquisition unit, used for splicing and convolving each of the first local space and spectral neighborhood features using a preset splicing and convolution rule to obtain a first splicing feature corresponding to each of the first local space and spectral neighborhood features;

[0043] a second stitching feature acquisition unit, configured to use a plurality of depth-separable convolution branches in the first local space and spectral feature compensation component to extract corresponding second local space and spectral neighborhood features from each of the first stitching features, and stitch and convolve each of the second local space and spectral neighborhood features using preset stitching and convolution rules to obtain second stitching features corresponding to each of the second local space and spectral neighborhood features;

[0044] The spectrum cluster and feature fusion unit is used to fuse each of the spectrum clusters to be processed with the corresponding second splicing feature using a preset residual connection to obtain a plurality of spectrum clusters to be reconstructed and a spectrum cluster set to be reconstructed including each of the spectrum clusters to be reconstructed.

[0045] Optionally, the second spectrum cluster set determination module includes:

[0046] A spectrum cluster conversion unit, used for performing a conversion operation on each of the spectrum clusters to be reconstructed in the set of spectrum clusters to be reconstructed in the preset non-local reconstruction feature extraction component, so as to obtain a corresponding initial two-dimensional sequence feature;

[0047] A sequence feature determination unit, used to process each of the initial two-dimensional sequence features in sequence using the first convolution layer, the root mean square normalization layer, and the linear transformation layer in the preset non-local reconstruction feature extraction component to obtain a corresponding two-dimensional sequence feature to be processed;

[0048] A first processing result acquisition unit is used to input each of the two-dimensional sequence features to be processed into the first branch, so as to use the second convolution layer, the first preset activation function and the preset visual state space model in the first branch to process each of the two-dimensional sequence features to be processed respectively to obtain a corresponding first processing result;

[0049] A second processing result acquisition unit is used to input each of the two-dimensional sequence features to be processed into the second branch, so as to process each of the two-dimensional sequence features to be processed respectively by using a second preset activation function in the second branch to obtain a corresponding second processing result;

[0050] The result inverse conversion unit is used to fuse the first processing result, the second processing result and the initial two-dimensional sequence feature by using the bottleneck residual connection in the preset non-local reconstruction feature extraction component to obtain a corresponding fusion result, and perform an inverse conversion operation on each of the fusion results to obtain a first reconstructed spectral cluster set.

[0051] As can be seen from the above, before performing anomaly detection of hyperspectral images based on the Mamba network model, the present application needs to use the preset sliding window spectral cluster expansion component and the first local space and spectral feature compensation component in the preset spectral cluster expansion component to process the original hyperspectral image in sequence to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image; use the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the spectral clusters to be reconstructed in the spectral cluster set to obtain a first reconstructed spectral cluster set; use the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component to process each of the first reconstructed spectral clusters in the first reconstructed spectral cluster set in sequence to obtain a second reconstructed spectral cluster set; and determine an anomaly detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectral cluster set.

[0052] It can be seen that the present application processes the original hyperspectral image in sequence by using the preset sliding window spectral cluster expansion component and the first local space and spectral feature compensation component to obtain a set of spectral clusters to be reconstructed; then. Use the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each spectral cluster to be reconstructed in the spectral cluster set to obtain a first reconstructed spectral cluster set; then. Use the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component to process each first reconstructed spectral cluster in the first reconstructed spectral cluster set in sequence to obtain a second reconstructed spectral cluster set; finally. Determine the anomaly detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectral cluster set. In this way, in the process of hyperspectral image anomaly detection based on the Mamba network model, the problem of detection performance degradation caused by limited non-local feature extraction and abnormal target contamination in the background reconstruction learning mode is prevented, thereby improving the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0054] Figure 1 A flow chart of a hyperspectral image anomaly detection method based on a Mamba network model disclosed in this application;

[0055] Figure 2A pseudo-color image and a distribution true value image of a specific hyperspectral image disclosed in this application;

[0056] Figure 3 A schematic diagram of the overall framework of a Mamba network model disclosed in this application;

[0057] Figure 4 A schematic diagram of the network structure of a first local space and spectral feature compensation component disclosed in the present application;

[0058] Figure 5 A schematic diagram of a network structure of a preset visual state space model disclosed in this application;

[0059] Figure 6 A schematic diagram of the structure of a visual 2D selective state space model disclosed in this application;

[0060] Figure 7 Schematic diagram of anomaly detection results corresponding to each anomaly detection method disclosed in this application;

[0061] Figure 8 This is a schematic diagram of the structure of a hyperspectral image anomaly detection system based on the Mamba network model disclosed in this application. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] At present, there are many hyperspectral anomaly detection algorithms on the market, such as anomaly detection methods based on data statistical distribution, based on collaborative and sparse representation models, based on low-rank sparse decomposition, and based on deep neural network models. However, the model-based method needs to optimize the model parameters according to the background and target characteristics of the hyperspectral image, and relies too much on manual features, which leads to limited performance of anomaly detection. In addition, the hyperspectral image anomaly detection method based on the Mamba network model based on the deep neural network model realizes anomaly detection by performing deep representation learning and difference feature extraction on the spectral sample data, but most of these methods have the problems of limited non-local feature extraction and high computational complexity, which limits the network detection performance of anomaly detection in the reconstruction learning mode. To this end, the present application provides a hyperspectral image anomaly detection method based on the Mamba network model, which can prevent the problem of detection performance degradation caused by limited non-local feature extraction and abnormal target contamination in the background reconstruction learning mode during the hyperspectral image anomaly detection process based on the Mamba network model.

[0064] See also Figure 1 As shown, an embodiment of the present invention discloses a hyperspectral image anomaly detection method based on a Mamba network model, which is applied to a Mamba network model including a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component, and a preset spectral cluster folding component. The method includes:

[0065] Step S11, using the preset sliding window spectral cluster expansion component and the first local space and spectral feature compensation component in the preset spectral cluster expansion component to process the original hyperspectral image in sequence, to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image.

[0066] In this embodiment, before detecting anomalies in hyperspectral images based on the Mamba network model, a preset sensor is used to collect the hyperspectral images that need to be detected for anomalies. In a specific embodiment, the preset sensor is a satellite-borne reflective optical system imaging spectrometer-03 (ROSIS-03) sensor, and the collected hyperspectral image is an image of a beach scene in Pavia, Italy. The spatial size of the image is 150×150 pixels, the number of bands is 188, and the above image contains seven abnormal vehicle targets. Figure 2 As shown, Figure 2 (a) in the figure is the pseudo-color image corresponding to the above image. Figure 2 (b) in the figure is the distribution truth map of the corresponding abnormal targets in the image.

[0067] It is worth mentioning that when the embodiment of the present application performs anomaly detection on the hyperspectral image, the Mamba (i.e., a selective structural state space model) network model is used to perform anomaly detection on the hyperspectral image, wherein the overall framework schematic diagram of the Mamba network model is as follows: Figure 3 As shown:

[0068] The Mamba network model includes a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component and a preset spectral cluster folding component, wherein the preset spectral cluster expansion component includes a preset sliding window spectral cluster expansion component, a first local space and spectral feature compensation component and a convolution fusion module; the preset non-local reconstruction feature extraction component includes a preset visual state space model; the preset spectral cluster folding component includes a convolution fusion module, a second local space and spectral feature compensation component and a preset sliding window spectral cluster folding component.

[0069] In this embodiment, the original hyperspectral image H needs to be input into the preset spectral cluster expansion component, so as to process the original hyperspectral image using the preset spectral cluster expansion component to obtain the spectral cluster set to be reconstructed. ,in, is the spectral cluster corresponding to the original hyperspectral image, N is the number of spectral clusters, and in addition, the preset spectral cluster expansion component is composed of a preset sliding window spectral cluster expansion component, a first local space and spectral feature compensation component and a convolution fusion module connected in series. Specifically, the preset sliding window spectral cluster expansion component and the first local space and spectral feature compensation component in the preset spectral cluster expansion component are used to process the original hyperspectral image in sequence to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image, which may include: using the preset sliding window spectral cluster expansion component in the preset spectral cluster expansion component to perform spectral cluster expansion on the original hyperspectral image to obtain a number of spectral clusters to be processed corresponding to the original hyperspectral image; the sizes of the spectral clusters to be processed are the same; in the first local space and spectral feature compensation component, a number of deeply separable convolution branches are used to extract features of each spectral cluster to be processed to obtain the corresponding first local space and spectral neighborhood features; the convolution kernel sizes corresponding to the deeply separable convolution branches are all different.

[0070] Furthermore, after obtaining the first local space and spectral neighborhood features corresponding to each spectral cluster to be processed, the embodiment of the present application first needs to use the preset splicing and convolution rules to splice and convolve each first local space and spectral neighborhood feature to obtain the first splicing features corresponding to each first local space and spectral neighborhood feature. Subsequently, in the first local space and spectral feature compensation component, a number of depth-separable convolution branches are used again to extract the corresponding second local space and spectral neighborhood features from each first splicing feature, and the preset splicing and convolution rules are used to splice and convolve each second local space and spectral neighborhood feature to obtain the second splicing features corresponding to each second local space and spectral neighborhood feature. Finally, the preset residual connection is used to fuse each spectral cluster to be processed with the corresponding second splicing feature to obtain a number of spectral clusters to be reconstructed and a set of spectral clusters to be reconstructed including each spectral cluster to be reconstructed.

[0071] In a specific embodiment, the preset sliding window spectral cluster expansion component divides the original hyperspectral image H into N partially overlapping spectral cluster sets in a left-to-right and top-to-bottom order through a preset sliding window. , where the length and width of the preset sliding window and the sliding step are set to 16 and 4 respectively, and the size of each spectral cluster is .

[0072] Furthermore, for each spectral cluster, the embodiment of the present application needs to input each spectral cluster into the first local space and spectral feature compensation component to extract multi-scale local space-spectral neighborhood features using multi-scale deep separable convolution, thereby compensating for the shortcoming of forgetting spatial local features in the linear scanning mechanism of the Mamba network model. Among them, the network structure of the first local space and spectral feature compensation component is as follows: Figure 4 As shown: for the spectral cluster input to the first local spatial and spectral feature compensation component, the embodiment of the present application will input it into three parallel depth-separable convolution branches respectively, where the convolution kernel sizes of the three branches are 3×3, 5×5 and 7×7 respectively, and then the following local multi-scale feature map is obtained through the GELU activation function:

[0073] ;

[0074] in, represents the GELU activation function, , as well as They represent the convolution kernel sizes respectively. , and Then, through channel concatenation and The convolution is used to fuse and interact the obtained multi-scale local features, and the deep separable convolution is used again to extract the fused and interacted features to compensate for the spatial-spectral information of the target local neighborhood. Finally, the residual connection is used to fuse with the original input spectral cluster to obtain the final output result as shown below:

[0075] ;

[0076] in, For channel splicing operations, is a 2D convolution with a kernel size of 1×1, Subsequently, for each spectral cluster after multi-scale spatial-spectral feature compensation, the embodiment of the present application performs channel integration through 1×1 2D convolution to uniformly adjust the number of channels to the number of spectral bands.

[0077] Step S12: using the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set, so as to obtain a first reconstructed spectrum cluster set.

[0078] In this embodiment, each spectral cluster after multi-scale spatial-spectral feature compensation needs to be sequentially The preset non-local reconstruction feature extraction component composed of a preset visual state space model is sent to perform background non-local reconstruction feature learning to obtain the first reconstructed spectral cluster set. Specifically, using a preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each to-be-reconstructed spectral cluster in the to-be-reconstructed spectral cluster set to obtain a first reconstructed spectral cluster set may include: performing conversion operations on each to-be-reconstructed spectral cluster in the to-be-reconstructed spectral cluster set in the preset non-local reconstruction feature extraction component to obtain corresponding initial two-dimensional sequence features; using a first convolution layer, a root mean square normalization layer, and a linear transformation layer in the preset non-local reconstruction feature extraction component to sequentially process each initial two-dimensional sequence feature to obtain corresponding two-dimensional sequence features to be processed; inputting each to-be-processed two-dimensional sequence feature into the first branch to obtain a first two-dimensional sequence feature. The second convolutional layer, the first preset activation function and the preset visual state space model in the first branch are used to process each of the two-dimensional sequence features to be processed respectively to obtain a corresponding first processing result; each of the two-dimensional sequence features to be processed is input into the second branch, so that the second preset activation function in the second branch is used to process each of the two-dimensional sequence features to be processed respectively to obtain a corresponding second processing result; the first processing result, the second processing result and the initial two-dimensional sequence feature are fused by the bottleneck residual connection in the preset non-local reconstruction feature extraction component to obtain a corresponding fusion result, and each fusion result is respectively inversely transformed to obtain a first reconstructed spectral cluster set.

[0079] Furthermore, the network structure of the preset visual state space model is as follows Figure 5 As shown: for the first reconstructed spectral cluster set input into the preset visual state space model, the embodiment of the present application first reshapes the first reconstructed spectral cluster set into a size of Two-dimensional sequence characteristics ,in, , , and then the two-dimensional sequence features are convolved by 1D convolution to convert its channel number Reduced to , and obtain the size of Two-dimensional sequence features.

[0080] Then, the embodiment of the present application needs to change the size to The two-dimensional sequence features are input into two parallel branches through the channel Chunk operation. as well as In one embodiment, the branch The output can be expressed as ;

[0081] in, as well as Representing 1D convolutional layer and visual 2D selective state space model respectively, Represents the SiLU activation function.

[0082] Further, another branch It consists of a single SiLU activation function, namely It is worth mentioning that the embodiment of the present application needs to connect the two branches and the two-dimensional sequence features through the bottleneck residual connection. The fusion is performed and the output result is The feature sequence is finally reshaped to output the two-dimensional sequence feature as a The three-dimensional feature map of the first reconstructed spectral cluster set .

[0083] Specifically, each two-dimensional sequence feature to be processed is input into the first branch, so that each two-dimensional sequence feature to be processed is processed separately by using the second convolution layer, the first preset activation function and the preset visual state space model in the first branch to obtain a corresponding first processing result, which may include: inputting each two-dimensional sequence feature to be processed into the first branch, so that each two-dimensional sequence feature to be processed is processed separately by using the second convolution layer and the first preset activation function in the first branch to obtain a corresponding feature to be fused; scanning each feature to be fused in the preset visual state space model by using a preset four-way scanning strategy, and fusing the results obtained by the scanning to obtain a corresponding first processing result; the first processing result is a non-local feature corresponding to each spectral cluster to be reconstructed.

[0084] It is worth mentioning that The structural diagram is as follows Figure 6 As shown: The core mechanism of the network is to adopt a preset four-way scanning strategy based on the linear scanning of the Mamba network model, that is, to scan from the four corners of the feature map to the relative positions and then fuse them: from the upper left to the lower right, from the lower right to the upper left, from the lower left to the upper right, and from the upper right to the lower left, to ensure that each element in the feature map integrates information from all other positions in different directions, thereby generating a global receptive field and achieving the extraction of non-local features without increasing the linear computational complexity.

[0085] Step S13: using the second local space and spectral feature compensation component in the preset spectral cluster folding component and the preset sliding window spectral cluster folding component to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain a second reconstructed spectral cluster set.

[0086] In this embodiment, the first reconstructed spectrum cluster set is uniformly input into the preset spectrum cluster folding component to obtain the second reconstructed spectrum cluster set after background reconstruction learning. , that is, the spectrum cluster set is fused to obtain a hyperspectral image of the original size. Wherein, the preset spectrum cluster folding component is composed of a serially connected convolution fusion module, a second local space and spectral feature compensation component and a preset sliding window spectrum cluster folding component. It is worth mentioning that the second local space and spectral feature compensation component and the convolution fusion module are similar in structure to the first local space and spectral feature compensation component, that is, the operation performed in the preset sliding window spectrum cluster folding component corresponds to the operation performed in the preset spectrum cluster expansion component. Specifically, the second local space and spectral feature compensation component and the preset sliding window spectrum cluster folding component in the preset spectrum cluster folding component are used to process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set in turn to obtain the second reconstructed spectrum cluster set, which may include: using the third convolution layer in the preset spectrum cluster folding component and the second local space and spectral feature compensation component to process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set in turn to obtain the corresponding compensated spectrum cluster; using the preset sliding window spectrum cluster folding component to perform folding operations on each compensated spectrum cluster to obtain the second reconstructed spectrum cluster set.

[0087] In this embodiment, the second reconstructed spectrum cluster set Spectral cluster collection with original hyperspectral Perform reconstruction loss calculation based on spectral clusters to optimize network parameters using the calculation results obtained. Specifically, after using the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain the second reconstructed spectral cluster set, it can also include: using the preset reconstruction loss calculation function to perform reconstruction loss calculation on the second reconstructed spectral cluster set and the spectral cluster set to be reconstructed to obtain the corresponding reconstruction loss result, and using the reconstruction loss result to adjust the model parameters of the Mamba network model.

[0088] In a specific implementation, the reconstruction loss calculation method is to use function, and the expression is as follows:

[0089] ;

[0090] in, The loss calculation result is: Represents a custom about The function of represents the learnable parameters of the network model, and N represents the number of spectral clusters.

[0091] Step S14: determining an anomaly detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectrum cluster set.

[0092] In this embodiment, after obtaining the network model after multiple iterations of training, the embodiment of the present application needs to save the network model after multiple iterations of training and convert the original hyperspectral image into Input it into the trained network model again to obtain the second reconstructed spectral cluster set , so as to obtain the reconstruction residual map corresponding to each spectral cluster using the preset reconstruction residual technology:

[0093] ;

[0094] in, as well as Respectively represent and The center is located at the coordinate pixel Specifically, determining the anomaly detection result map corresponding to the original hyperspectral image based on the second reconstructed spectrum cluster set may include: using a preset reconstruction residual technology and determining the reconstruction residual map corresponding to each reconstructed spectrum cluster in the second reconstructed spectrum cluster set based on the second reconstructed spectrum cluster set. Finally, using a preset fusion technology to fuse each reconstruction residual map to obtain the anomaly detection result map corresponding to the original hyperspectral image.

[0095] It is worth mentioning that Figure 7 (a), (b), (c) and (d) are the anomaly detection result diagrams obtained by performing anomaly detection on the Pavia hyperspectral image using the PCA-TLRSR anomaly detection method, the KIFD anomaly detection method, the DFAN anomaly detection method based on a deep neural network and the anomaly detection method proposed in the embodiment of the present application. Figure 7 It can be seen that compared with the PCA-TLRSR anomaly detection method, the KIFD anomaly detection method and the DFAN anomaly detection method based on deep neural network, the anomaly detection method of the present invention achieves better background suppression and false alarm suppression, so that the abnormal target and the background can be more significantly separated, while the other three anomaly detection methods have more false alarms, resulting in the inability to distinguish between the background and the anomaly.

[0096] As can be seen from the above, before the hyperspectral image anomaly detection based on the Mamba network model is performed in the embodiment of the present application, the preset sliding window spectral cluster expansion component and the first local space and spectral feature compensation component in the preset spectral cluster expansion component are used to process the original hyperspectral image in sequence to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image; the preset non-local reconstruction feature extraction component is used to perform non-local feature extraction and background reconstruction feature learning on each spectral cluster to be reconstructed in the spectral cluster set to be reconstructed, so as to obtain a first reconstructed spectral cluster set; the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component are used to process each first reconstructed spectral cluster in the first reconstructed spectral cluster set in sequence to obtain a second reconstructed spectral cluster set; based on the second reconstructed spectral cluster set, the anomaly detection result graph corresponding to the original hyperspectral image is determined. In this way, in the process of hyperspectral image anomaly detection based on the Mamba network model, the problem of detection performance degradation caused by limited non-local feature extraction and abnormal target contamination in the background reconstruction learning mode is prevented.

[0097] Accordingly, see Figure 8 As shown, the present application also provides a hyperspectral image anomaly detection system based on a Mamba network model, which is applied to a Mamba network model including a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component, and a preset spectral cluster folding component, and the system includes:

[0098] A first spectrum cluster set determination module 11 is used to sequentially process the original hyperspectral image using a preset sliding window spectrum cluster expansion component and a first local space and spectrum feature compensation component in the preset spectrum cluster expansion component to obtain a spectrum cluster set to be reconstructed corresponding to the original hyperspectral image;

[0099] The second spectrum cluster set determination module 12 is used to use the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the spectrum clusters to be reconstructed in the spectrum cluster set to be reconstructed, so as to obtain a first reconstructed spectrum cluster set;

[0100] A third spectrum cluster set determination module 13 is used to sequentially process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set by using the second local space and spectrum feature compensation component and the preset sliding window spectrum cluster folding component in the preset spectrum cluster folding component to obtain a second reconstructed spectrum cluster set;

[0101] The detection result graph determining module 14 is configured to determine an abnormality detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectrum cluster set.

[0102] As can be seen from the above, before performing anomaly detection of hyperspectral images based on the Mamba network model, the embodiment of the present application needs to process the original hyperspectral image in sequence by using the preset sliding window spectral cluster expansion component and the first local space and spectral feature compensation component to obtain a set of spectral clusters to be reconstructed; then. Use the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each spectral cluster to be reconstructed in the spectral cluster set to obtain a first reconstructed spectral cluster set; then. Use the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component to process each first reconstructed spectral cluster in the first reconstructed spectral cluster set in sequence to obtain a second reconstructed spectral cluster set; finally. Determine the anomaly detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectral cluster set. In this way, in the process of hyperspectral image anomaly detection based on the Mamba network model, the problem of detection performance degradation caused by limited non-local feature extraction and abnormal target contamination in the background reconstruction learning mode is prevented.

[0103] In some specific implementations, the first spectrum cluster set determination module 11 may specifically include:

[0104] a spectral cluster expansion unit, configured to perform spectral cluster expansion on the original hyperspectral image using a preset sliding window spectral cluster expansion component in the preset spectral cluster expansion component, so as to obtain a plurality of spectral clusters to be processed corresponding to the original hyperspectral image; the sizes of the spectral clusters to be processed are the same;

[0105] A feature extraction unit, used for performing feature extraction on each of the spectral clusters to be processed by using a plurality of depth-separable convolution branches in the first local space and spectral feature compensation component to obtain corresponding first local space and spectral neighborhood features; the convolution kernel sizes corresponding to the depth-separable convolution branches are different;

[0106] A first splicing feature acquisition unit, used for splicing and convolving each of the first local space and spectral neighborhood features using a preset splicing and convolution rule to obtain a first splicing feature corresponding to each of the first local space and spectral neighborhood features;

[0107] a second stitching feature acquisition unit, configured to use a plurality of depth-separable convolution branches in the first local space and spectral feature compensation component to extract corresponding second local space and spectral neighborhood features from each of the first stitching features, and stitch and convolve each of the second local space and spectral neighborhood features using preset stitching and convolution rules to obtain second stitching features corresponding to each of the second local space and spectral neighborhood features;

[0108] The spectrum cluster and feature fusion unit is used to fuse each of the spectrum clusters to be processed with the corresponding second splicing feature using a preset residual connection to obtain a plurality of spectrum clusters to be reconstructed and a spectrum cluster set to be reconstructed including each of the spectrum clusters to be reconstructed.

[0109] In some specific implementations, the second spectrum cluster set determination module 12 may specifically include:

[0110] A spectrum cluster conversion unit, used for performing a conversion operation on each of the spectrum clusters to be reconstructed in the set of spectrum clusters to be reconstructed in the preset non-local reconstruction feature extraction component, so as to obtain a corresponding initial two-dimensional sequence feature;

[0111] A sequence feature determination unit, used to process each of the initial two-dimensional sequence features in sequence using the first convolution layer, the root mean square normalization layer, and the linear transformation layer in the preset non-local reconstruction feature extraction component to obtain a corresponding two-dimensional sequence feature to be processed;

[0112] A first processing result acquisition unit is used to input each of the two-dimensional sequence features to be processed into the first branch, so as to use the second convolution layer, the first preset activation function and the preset visual state space model in the first branch to process each of the two-dimensional sequence features to be processed respectively to obtain a corresponding first processing result;

[0113] A second processing result acquisition unit is used to input each of the two-dimensional sequence features to be processed into the second branch, so as to process each of the two-dimensional sequence features to be processed respectively by using a second preset activation function in the second branch to obtain a corresponding second processing result;

[0114] The result inverse conversion unit is used to fuse the first processing result, the second processing result and the initial two-dimensional sequence feature by using the bottleneck residual connection in the preset non-local reconstruction feature extraction component to obtain a corresponding fusion result, and perform an inverse conversion operation on each of the fusion results to obtain a first reconstructed spectral cluster set.

[0115] Accordingly, in some specific implementations, the second spectrum cluster set determination module 12 may specifically include:

[0116] A feature determination unit to be fused, used for inputting each of the two-dimensional sequence features to be processed into the first branch, so as to process each of the two-dimensional sequence features to be processed respectively by using the second convolution layer and the first preset activation function in the first branch to obtain the corresponding feature to be fused;

[0117] A feature scanning unit is used to scan each of the features to be fused in the preset visual state space model using a preset four-way scanning strategy, and fuse the scanning results to obtain a corresponding first processing result; the first processing result is a non-local feature corresponding to each of the spectral clusters to be reconstructed.

[0118] In some specific implementations, the third spectrum cluster set determination module 13 may specifically include:

[0119] a spectral cluster processing unit, configured to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set by using the third convolution layer and the second local space and spectral feature compensation component in the preset spectral cluster folding component, so as to obtain a corresponding compensated spectral cluster;

[0120] The spectrum cluster folding unit is used to perform folding operations on each of the compensated spectrum clusters using a preset sliding window spectrum cluster folding component to obtain a second reconstructed spectrum cluster set.

[0121] In some specific implementations, the hyperspectral image anomaly detection system based on the Mamba network model may further include:

[0122] The model parameter adjustment unit is used to calculate the reconstruction loss of the second reconstructed spectrum cluster set and the to-be-reconstructed spectrum cluster set using a preset reconstruction loss calculation function to obtain corresponding reconstruction loss results, so as to adjust the model parameters of the Mamba network model using the reconstruction loss results.

[0123] In some specific implementations, the detection result graph determination module 14 may specifically include:

[0124] a residual map determining unit, configured to determine, by using a preset reconstruction residual technique and based on the second reconstructed spectrum cluster set, a reconstruction residual map respectively corresponding to each of the reconstructed spectrum clusters in the second reconstructed spectrum cluster set;

[0125] The residual image fusion unit is used to fuse the reconstructed residual images using a preset fusion technology to obtain an anomaly detection result image corresponding to the original hyperspectral image.

[0126] Furthermore, the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0127] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0128] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0129] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0130] The technical solution provided by the present application is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technicians in this field, according to the idea of ​​the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A hyperspectral image anomaly detection method based on the Mamba network model, characterized in that: Applied to a Mamba network model including a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component and a preset spectral cluster folding component, the method comprises: The original hyperspectral image is processed in sequence by using a preset sliding window spectral cluster expansion component and a first local space and spectral feature compensation component in the preset spectral cluster expansion component to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image; Using the preset non-local reconstruction feature extraction component, respectively perform non-local feature extraction and background reconstruction feature learning on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set, so as to obtain a first reconstructed spectrum cluster set; Using the second local space and spectral feature compensation component in the preset spectral cluster folding component and the preset sliding window spectral cluster folding component to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain a second reconstructed spectral cluster set; An anomaly detection result map corresponding to the original hyperspectral image is determined based on the second reconstructed spectrum cluster set.

2. The hyperspectral image anomaly detection method based on the Mamba network model according to claim 1 is characterized in that: The method of using a preset sliding window spectral cluster expansion component and a first local space and spectral feature compensation component in the preset spectral cluster expansion component to sequentially process the original hyperspectral image to obtain a set of spectral clusters to be reconstructed corresponding to the original hyperspectral image includes: Utilizing a preset sliding window spectral cluster expansion component in the preset spectral cluster expansion component to perform spectral cluster expansion on the original hyperspectral image, so as to obtain a plurality of spectral clusters to be processed corresponding to the original hyperspectral image; the sizes of the spectral clusters to be processed are the same; In the first local space and spectral feature compensation component, a plurality of depth-separable convolution branches are used to extract features of each of the spectral clusters to be processed, so as to obtain corresponding first local space and spectral neighborhood features; the convolution kernel sizes corresponding to the depth-separable convolution branches are different; Using preset splicing and convolution rules, each of the first local space and spectral neighborhood features is spliced ​​and convolved to obtain first splicing features corresponding to each of the first local space and spectral neighborhood features; In the first local space and spectral feature compensation component, a plurality of depth-separable convolution branches are again used to extract corresponding second local space and spectral neighborhood features from each of the first spliced ​​features, and each of the second local space and spectral neighborhood features is spliced ​​and convolved using a preset splicing and convolution rule to obtain second splicing features corresponding to each of the second local space and spectral neighborhood features; Each of the to-be-processed spectrum clusters is fused with the corresponding second splicing feature by using a preset residual connection to obtain a plurality of to-be-reconstructed spectrum clusters and a to-be-reconstructed spectrum cluster set including each of the to-be-reconstructed spectrum clusters.

3. The hyperspectral image anomaly detection method based on the Mamba network model according to claim 1 is characterized in that: The method of using the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set to obtain a first reconstructed spectrum cluster set includes: In the preset non-local reconstruction feature extraction component, a conversion operation is performed on each of the to-be-reconstructed spectrum clusters in the to-be-reconstructed spectrum cluster set to obtain a corresponding initial two-dimensional sequence feature; Using the first convolution layer, the root mean square normalization layer and the linear transformation layer in the preset non-local reconstruction feature extraction component to process each of the initial two-dimensional sequence features in turn, to obtain corresponding two-dimensional sequence features to be processed; Inputting each of the two-dimensional sequence features to be processed into the first branch, so as to process each of the two-dimensional sequence features to be processed respectively by using the second convolutional layer, the first preset activation function and the preset visual state space model in the first branch to obtain a corresponding first processing result; Inputting each of the two-dimensional sequence features to be processed into the second branch, so as to process each of the two-dimensional sequence features to be processed respectively using a second preset activation function in the second branch to obtain a corresponding second processing result; The first processing result, the second processing result and the initial two-dimensional sequence feature are fused by using the bottleneck residual connection in the preset non-local reconstruction feature extraction component to obtain a corresponding fusion result, and each of the fusion results is respectively subjected to an inverse transformation operation to obtain a first reconstructed spectral cluster set.

4. The hyperspectral image anomaly detection method based on the Mamba network model according to claim 3 is characterized in that: The step of inputting each of the two-dimensional sequence features to be processed into the first branch so as to process each of the two-dimensional sequence features to be processed respectively by using the second convolution layer, the first preset activation function and the preset visual state space model in the first branch to obtain a corresponding first processing result includes: Inputting each of the two-dimensional sequence features to be processed into the first branch, so as to use the second convolution layer and the first preset activation function in the first branch to process each of the two-dimensional sequence features to be processed respectively, so as to obtain corresponding features to be fused; In the preset visual state space model, each of the features to be fused is scanned using a preset four-way scanning strategy, and the scanned results are fused to obtain a corresponding first processing result; the first processing result is a non-local feature corresponding to each of the spectral clusters to be reconstructed.

5. The hyperspectral image anomaly detection method based on the Mamba network model according to claim 1 is characterized in that: The method of using the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain a second reconstructed spectral cluster set includes: Using the third convolution layer in the preset spectrum cluster folding component and the second local space and spectrum feature compensation component to sequentially process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set, so as to obtain a corresponding compensated spectrum cluster; A preset sliding window spectrum cluster folding component is used to perform a folding operation on each of the compensated spectrum clusters to obtain a second reconstructed spectrum cluster set.

6. The hyperspectral image anomaly detection method based on the Mamba network model according to claim 1 is characterized in that: After the second local space and spectral feature compensation component and the preset sliding window spectral cluster folding component in the preset spectral cluster folding component are used to sequentially process each first reconstructed spectral cluster in the first reconstructed spectral cluster set to obtain a second reconstructed spectral cluster set, the method further includes: A preset reconstruction loss calculation function is used to perform reconstruction loss calculation on the second reconstructed spectrum cluster set and the to-be-reconstructed spectrum cluster set to obtain corresponding reconstruction loss results, so as to use the reconstruction loss results to adjust the model parameters of the Mamba network model.

7. The hyperspectral image anomaly detection method based on the Mamba network model according to any one of claims 1 to 6, characterized in that: The determining, based on the second reconstructed spectrum cluster set, an anomaly detection result graph corresponding to the original hyperspectral image comprises: Determine, by using a preset reconstruction residual technique and based on the second reconstructed spectrum cluster set, a reconstruction residual graph corresponding to each of the reconstructed spectrum clusters in the second reconstructed spectrum cluster set; The reconstructed residual images are fused using a preset fusion technology to obtain an anomaly detection result image corresponding to the original hyperspectral image.

8. A hyperspectral image anomaly detection system based on the Mamba network model, characterized in that: The system is applied to a Mamba network model including a preset spectral cluster expansion component, a preset non-local reconstruction feature extraction component and a preset spectral cluster folding component, and the system includes: A first spectrum cluster set determination module is used to sequentially process the original hyperspectral image using a preset sliding window spectrum cluster expansion component and a first local space and spectrum feature compensation component in the preset spectrum cluster expansion component to obtain a spectrum cluster set to be reconstructed corresponding to the original hyperspectral image; A second spectrum cluster set determination module is used to use the preset non-local reconstruction feature extraction component to perform non-local feature extraction and background reconstruction feature learning on each of the spectrum clusters to be reconstructed in the spectrum cluster set to be reconstructed, so as to obtain a first reconstructed spectrum cluster set; a third spectrum cluster set determination module, configured to sequentially process each first reconstructed spectrum cluster in the first reconstructed spectrum cluster set by using the second local space and spectrum feature compensation component and the preset sliding window spectrum cluster folding component in the preset spectrum cluster folding component to obtain a second reconstructed spectrum cluster set; The detection result graph determining module is used to determine an abnormality detection result graph corresponding to the original hyperspectral image based on the second reconstructed spectrum cluster set.

9. The hyperspectral image anomaly detection system based on the Mamba network model according to claim 8, characterized in that: The first spectrum cluster set determination module includes: a spectral cluster expansion unit, configured to perform spectral cluster expansion on the original hyperspectral image using a preset sliding window spectral cluster expansion component in the preset spectral cluster expansion component, so as to obtain a plurality of spectral clusters to be processed corresponding to the original hyperspectral image; the sizes of the spectral clusters to be processed are the same; A feature extraction unit, used for performing feature extraction on each of the spectral clusters to be processed by using a plurality of depth-separable convolution branches in the first local space and spectral feature compensation component to obtain corresponding first local space and spectral neighborhood features; the convolution kernel sizes corresponding to the depth-separable convolution branches are different; A first splicing feature acquisition unit, used for splicing and convolving each of the first local space and spectral neighborhood features using a preset splicing and convolution rule to obtain a first splicing feature corresponding to each of the first local space and spectral neighborhood features; a second stitching feature acquisition unit, configured to use a plurality of depth-separable convolution branches in the first local space and spectral feature compensation component to extract corresponding second local space and spectral neighborhood features from each of the first stitching features, and stitch and convolve each of the second local space and spectral neighborhood features using preset stitching and convolution rules to obtain second stitching features corresponding to each of the second local space and spectral neighborhood features; The spectrum cluster and feature fusion unit is used to fuse each of the spectrum clusters to be processed with the corresponding second splicing feature using a preset residual connection to obtain a plurality of spectrum clusters to be reconstructed and a spectrum cluster set to be reconstructed including each of the spectrum clusters to be reconstructed.

10. The hyperspectral image anomaly detection system based on the Mamba network model according to claim 8, characterized in that: The second spectrum cluster set determination module includes: A spectrum cluster conversion unit, used for performing a conversion operation on each of the spectrum clusters to be reconstructed in the set of spectrum clusters to be reconstructed in the preset non-local reconstruction feature extraction component, so as to obtain a corresponding initial two-dimensional sequence feature; A sequence feature determination unit, used to process each of the initial two-dimensional sequence features in sequence using the first convolution layer, the root mean square normalization layer, and the linear transformation layer in the preset non-local reconstruction feature extraction component to obtain a corresponding two-dimensional sequence feature to be processed; A first processing result acquisition unit is used to input each of the two-dimensional sequence features to be processed into the first branch, so as to use the second convolution layer, the first preset activation function and the preset visual state space model in the first branch to process each of the two-dimensional sequence features to be processed respectively to obtain a corresponding first processing result; A second processing result acquisition unit is used to input each of the two-dimensional sequence features to be processed into the second branch, so as to process each of the two-dimensional sequence features to be processed respectively by using a second preset activation function in the second branch to obtain a corresponding second processing result; The result inverse conversion unit is used to fuse the first processing result, the second processing result and the initial two-dimensional sequence feature by using the bottleneck residual connection in the preset non-local reconstruction feature extraction component to obtain a corresponding fusion result, and perform an inverse conversion operation on each of the fusion results to obtain a first reconstructed spectral cluster set.

Images