Hyperspectral anomaly detection method based on background clustering constraint under potential feature separation

Abstract

The application discloses a hyperspectral anomaly detection method based on background clustering constraint and potential feature separation, and relates to the technical field of hyperspectral remote sensing detection applications. First, band selection and partition generation of a hyperspectral original image are carried out to generate fixed-size image blocks. Then, a basic image reconstruction framework is constructed based on a lightweight spectral-spatial feature extraction encoder and a convolution decoder. Then, a pixel-level deep clustering network is introduced to complete latent feature clustering modeling, calibrate the background clustering label of the image block, and obtain the intra-class RX detection score in the feature space. Then, the intra-class RX score and the reconstructed image are reversely mapped to the original image space to construct a global RX response map and a reconstructed full image. Finally, a weight matrix is generated by using the reconstruction error between the original image and the reconstructed full image, the global RX response map is modulated, and the final anomaly detection result is output. The application fuses the background clustering constraint and the potential feature separation mechanism, effectively improves the detection accuracy of small anomaly targets, and significantly suppresses false alarm interference.

Description

Technical Field

[0001] This application relates to the field of hyperspectral remote sensing application technology, and in particular to a hyperspectral anomaly detection method based on latent feature separation under background clustering constraints. Background Technology

[0002] Spaceborne hyperspectral imaging technology acquires three-dimensional "geometric-radiometric-spectral" information of targets simultaneously across dozens to hundreds of consecutive narrow bands, achieving precise "image-spectrum integration" detection. It has been widely applied in fields such as mineral resource exploration, ecological environment monitoring, disaster early warning, and national defense. Hyperspectral anomaly detection, as its core application technology, aims to identify targets that occupy an extremely low percentage of pixels in the spatial dimension (0.0019%-1% of the entire image) and exhibit a spectral response significantly different from the surrounding background. Compared to traditional hyperspectral target identification, which relies on complete prior spectral information, hyperspectral anomaly detection can operate stably even with missing or insufficient prior information, possessing stronger adaptability to complex scenes and a wider range of applications.

[0003] In existing technologies, based on the different strategies used to distinguish between background and anomalies, hyperspectral anomaly detection can be mainly divided into three categories:

[0004] (1) Statistical methods: The RX algorithm (proposed by Reed and Xiaoli) is a classic statistical anomaly detection method. It assumes that the background follows a multivariate Gaussian distribution and relies on Mahalanobis distance to identify anomalies. Subsequent improved algorithms such as CRD optimize the adaptability of non-uniform scenes through local windows. However, this type of method is limited by the assumption of a single distribution. In complex multimodal backgrounds, the modeling accuracy is low and the false alarm rate is high. It cannot effectively characterize high-dimensional spectral nonlinear correlations.

[0005] (2) Representation-based methods: Based on the prior of "low-rank background and sparse anomalies", the image is decomposed to separate the background from the anomalies. The representative algorithms are robust and resistant to noise. However, the global low-rank assumption is not applicable to non-stationary spatial backgrounds, which can easily lead to missed anomaly detection. In addition, the lack of background category constraints results in insufficient feature mining under complex background interference, and the effect of distinguishing between anomalies and background is limited.

[0006] (3) Deep learning-based methods: These methods automatically mine joint spatial and spectral features using various deep models and rely on reconstruction errors to complete anomaly identification, resulting in strong model adaptability. However, they have shortcomings such as dependence on labeled data, high complexity, and weak interpretability, making it difficult to adapt to spaceborne computing power and bandwidth limitations. At the same time, they generally ignore the multi-cluster characteristics of the background, lack the ability to represent multiple backgrounds, and are prone to background overfitting and anomaly underfitting problems.

[0007] In summary, existing hyperspectral anomaly detection technologies are ill-suited for complex spaceborne applications, exhibiting significant technical bottlenecks. Traditional global statistics and low-rank modeling struggle to adapt to multimodal, non-stationary surface backgrounds, easily leading to false alarms and missed detections. Deep learning algorithms rely heavily on raw, shallow spectral features, lacking effective separation and enhancement of local spatial features. Insufficient fusion of spatial and spectral information makes it difficult to suppress mixed pixels and noise interference. Furthermore, the overall performance of these algorithms is unbalanced, failing to meet the detection requirements of high-dimensional, low signal-to-noise ratio, and small targets in spaceborne data. Especially in large-scale, complex scenarios, strong background heterogeneity interference makes it easy for small anomalies to be masked, limiting practical applications. Summary of the Invention

[0008] To address the above technical problems, this invention provides a hyperspectral anomaly detection method based on latent feature separation under background clustering constraints. This method aims to solve the problem that existing technologies often fail to detect small anomalies in complex scenes due to inaccurate background modeling, insufficient feature mining, and high false alarm rates.

[0009] Therefore, this application adopts the following technical solution:

[0010] A hyperspectral anomaly detection method based on latent feature separation under background clustering constraints includes:

[0011] Step 1: Perform band selection processing on the original hyperspectral image data and generate fixed-size image blocks through spatial partitioning;

[0012] Step 2: Construct a hyperspectral anomaly detection network. Use a lightweight spectral-spatial feature extractor to build an encoder to mine global semantic information and deep spectral-spatial features of image patches. Introduce a skip connection mechanism to pass the stitched spectral-spatial features to the decoder. Use a convolutional mapper to build a decoder to generate the reconstructed image.

[0013] Step 3: A pixel-level deep clustering network is introduced between the encoder and decoder. A Gaussian mixture model is applied to the latent spectral-spatial features extracted by the pixel-level deep clustering network to perform latent constraints. The latent feature clustering model is completed based on the Gaussian probability density loss function, thereby labeling the background clustering labels of the image blocks. For the labeled background clustering labels of the image blocks, the RX detection algorithm is applied to generate the intra-class RX detection scores in the feature space.

[0014] Step 4: The intra-class RX detection scores in the feature space are mapped back to the original hyperspectral image using the aggregation operator to form a global RX anomaly response map and a reconstructed full image.

[0015] Step 5: Perform decision fusion on the global RX anomaly response map and the reconstructed full map to form the final hyperspectral image anomaly detection result map.

[0016] This invention offers the following advantages: It proposes a "background clustering constraint + latent feature separation" technical approach. A lightweight spectral-spatial feature extractor and a convolutional mapper are used to construct an anomaly detection network. Latent feature separation effectively removes redundant background information and anomalous target features, enhancing the representational power of latent spectral-spatial features. Simultaneously, a pixel-level deep clustering network is embedded between the encoder and decoder. A Gaussian mixture model is used to constrain latent features. By clustering and modeling the hyperspectral background, the multi-class structure of the background is explicitly constrained, significantly reducing the interference of complex background heterogeneity on anomaly detection and solving the problem of inaccurate background modeling in existing methods. Intra-class RX detection in a low-variance feature space replaces global RX, avoiding detection bias caused by background heterogeneity and adapting to large-scale complex surface scenes. Finally, a weight matrix is ​​constructed based on image reconstruction errors, and the global anomaly response map is adaptively weighted and modulated, effectively improving the detection accuracy of small anomalies and significantly reducing the probability of false alarms and missed alarms in complex scenes. It is suitable for the real-world scenario of "scarcity of anomalous samples" in the field of hyperspectral anomaly detection. Moreover, the overall model has a small number of parameters and high computational efficiency, which can adapt to the computing power and bandwidth limitations of spaceborne platforms and better meet the detection needs of weak hyperspectral anomalous targets in a wide range of complex scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a hyperspectral anomaly detection method based on latent feature separation under background clustering constraints, provided in an embodiment of this application.

[0019] Figure 2 This is a framework diagram of a hyperspectral anomaly detection method based on latent feature separation under background clustering constraints, provided in an embodiment of this application.

[0020] Figure 3 This application provides a visualization of the correlation coefficient and band index between adjacent bands in a hyperspectral anomaly detection method based on latent feature separation under background clustering constraints.

[0021] Figure 4 This is a diagram of an encoder architecture based on a lightweight spectral-spatial feature extractor provided in an embodiment of this application.

[0022] Figure 5 A pixel-level deep clustering network architecture diagram provided for embodiments of this application.

[0023] Figure 6 The method provided in this application embodiment is compared with the color anomaly maps obtained by various hyperspectral anomaly detection methods in the prior art on the GF5-AHSI experimental dataset.

[0024] Figure 7 The ROC curves of the method provided in this application embodiment are compared with those of various existing hyperspectral anomaly detection methods on the GF5-AHSI experimental dataset.

[0025] Figure 8 The box-separable plots of anomaly detection on the GF5-AHSI experimental dataset are compared between the method provided in this application embodiment and various existing hyperspectral anomaly detection methods. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0027] The following detailed description, in conjunction with the accompanying drawings, describes a hyperspectral anomaly detection method based on latent feature separation under background clustering constraints, according to an embodiment of this application. It proposes a "background clustering constraint + latent feature separation" technical approach, which mines latent spectral-spatial joint features through a lightweight encoding / decoding structure, achieves intra-class anomaly detection by relying on deep clustering to model background clusters, and combines adaptive weighting of reconstruction errors. This effectively suppresses interference from complex backgrounds, improves the detection accuracy of weak anomalies, reduces the probability of false alarms and missed detections, adapts to the computing power and bandwidth limitations of spaceborne platforms, and better meets the practical application needs of hyperspectral anomaly detection in large-scale complex scenarios.

[0028] refer to Figure 1 and Figure 2 The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints provided in this application includes the following steps:

[0029] In step S101, the hyperspectral raw image data is subjected to band selection processing, and fixed-size image blocks are generated through spatial partitioning;

[0030] In step S102, a hyperspectral anomaly detection network is constructed, and a lightweight spectral-spatial feature extractor is used to build an encoder to efficiently mine global semantic information and deep spectral-spatial features of image patches; a skip connection mechanism is introduced to transmit the stitched spectral-spatial features to the decoder; and a convolutional mapper is used to build a decoder to generate the reconstructed image.

[0031] In step S103, a pixel-level deep clustering network is introduced between the encoder and the decoder. A Gaussian mixture model is applied to the extracted latent spectral-spatial features to perform latent constraints. The latent feature clustering model is completed based on the Gaussian probability density loss function, thereby labeling the background clustering labels of the image blocks. For the labeled background clustering labels of the image blocks, the RX detection algorithm is applied to generate the intra-class RX detection scores in the feature space.

[0032] In step S104, the RX score within the feature space and the reconstructed image are mapped back to the original hyperspectral image by the aggregation operator to form a global RX anomaly response map and a reconstructed full image;

[0033] In step S105, the global RX anomaly response map and the reconstructed full map are fused by decision to form the final hyperspectral image anomaly detection result map.

[0034] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0035] In some embodiments of this application, the band selection process involves calculating the correlation coefficients of adjacent bands in the hyperspectral raw image data to determine the correlation distribution of the hyperspectral raw image data; adaptively dividing the hyperspectral raw image data into several data subspaces based on the minimum value of the correlation coefficients of adjacent bands and a set threshold; calculating the band indices of all bands in each data subspace; and selecting a preset number of bands with the largest band indices as representative bands in each data subspace in the same proportion; arranging the representative bands selected from each data subspace in the initial band order to form a hyperspectral feature map after band selection. ,in Let be the set of real numbers. and These represent the length and width of the original hyperspectral image, respectively. Select the number of bands after band selection.

[0036] Furthermore, the correlation coefficients between adjacent bands of the original hyperspectral image data are:

[0037] ;

[0038] The indexes for all bands within each data subspace are:

[0039] ;

[0040] in, For the first The band and the first Correlation coefficient between adjacent bands For the first The band and the first The correlation coefficients between adjacent bands; N is the total number of samples in the original hyperspectral image data. The pixel ordinal number. For the first The first band image within the [number] band pixel value, For the first The first band image within the [number] band Each pixel value; For the first The average value of all pixels in each band. For the first The average value of all pixels in each band; where For the first The standard deviation of all pixels in each band.

[0041] Band selection can reduce data dimensionality while preserving the physical meaning and spectral characteristics of the data, which is of great help in improving anomaly detection efficiency. See also Figure 3 By calculating the correlation coefficients of adjacent bands in hyperspectral data It is evident that the original data exhibits strong correlation and segmentation effects. The larger the value, the richer the information within that band, and the weaker the correlation between bands. Band selection processing can ensure the preservation of certain diagnostic features related to band order.

[0042] In some embodiments of this application, fixed-size image patches are generated through spatial partitioning. Further, a window of a certain size is used to slide across the hyperspectral feature map after band processing with a certain sliding step size to extract overlapping three-dimensional neighborhood image patches, generating images from these patches. A dataset consisting of three-dimensional domain image patches.

[0043] For example, using a size of The window, with sliding steps Slide the hyperspectral feature map S after band selection to extract overlapping 3D neighborhood image patches, generating a dataset containing... A set of three-dimensional neighborhood image patches P∈ , of which The representation of an image patch is as follows ;in, Let be the set of real numbers. Select the number of bands after band selection. , They represent the first The first pixel in the upper left corner of each image patch corresponds to the hyperspectral feature map after band selection. Rows and columns in This represents the total number of image patches.

[0044] In some embodiments of this application, the process of constructing an encoder using a lightweight spectral-spatial feature extractor to efficiently mine global semantic information and deep spectral-spatial features of image patches is as follows: a three-dimensional convolutional module is used to extract joint features of the spectral and spatial dimensions of the image patch, outputting a local spectral-spatial feature map; subsequently, a two-dimensional convolutional module is introduced to extract features again from the latent spectral-spatial feature map output by the three-dimensional convolutional module, generating a feature map after spatial dimension reconstruction and enhancement. Next, based on the Transformer architecture, the reconstructed and enhanced feature maps with spatial dimensions are normalized and input into a multi-head self-attention mechanism to establish long-distance dependencies and generate global spectral-spatial feature maps. Finally, a lightweight bottleneck convolutional network is constructed to enhance the local spatial modeling capability of the feature maps. Through the above steps, efficient mining of global semantic information and deep spectral-spatial features of image patches is achieved.

[0045] Furthermore, such as Figure 4 As shown, the 3D convolutional module includes a 3D convolutional layer (CNN), a normalization layer (BN), and an activation function (ReLU) to capture... The spectral-spatial joint information is used to output a deep spectral-spatial feature map. ∈ ,in, This refers to the number of 3D neighborhood image patches generated after 3D convolution. , and These represent the height, width, and depth of the feature block after passing through the 3D convolutional layer, respectively. This effectively captures 3D correlation while avoiding feature loss and computational redundancy caused by deep networks. The formula is shown below:

[0046] ;

[0047] in, and These are the 3D convolution kernel weights and biases, respectively. For normalization function, This represents the standard convolution operation.

[0048] Furthermore, such as Figure 4 As shown, the two-dimensional convolutional module includes a two-dimensional depthwise separable convolutional layer (DSC), a normalization layer (BN), and an activation function (ReLU). Fine-grained spectral-spatial features resulting from the merging of channel and depth dimensions Then, feature extraction is performed to generate spatially reconstructed and enhanced feature maps. To further enhance local spatial features and control computational costs, the formula is as follows:

[0049] ;

[0050] in, For normalization function, This indicates independent convolution by channel. and These are the sets of weights for depthwise convolutional kernels and pointwise convolutional kernels, respectively. This represents the pointwise convolution bias vector.

[0051] While the Spectral-Spatial Encoding Labeller (SSET) composed of the aforementioned combination of 3D convolution and 2D depthwise separable convolution reduces the difficulty and computational cost of Transformer modeling, traditional Transformers struggle to extract correlations between neighboring pixels while modeling global spectral dependencies. Therefore, this application introduces a lightweight feedforward bottleneck convolutional network (FBCN) into the feedforward neural network multilayer perceptron (MLP) structure to enhance the local spatial modeling capability of feature maps.

[0052] Specifically, the output sequence of the multi-head self-attention (MHSA) in the Transformer architecture is passed through (Sequence to 2D Feature Map Reconstruction Operation) Reconstructing into a 2D feature map To introduce locality:

[0053] ;

[0054] in, This is the output sequence of multi-head self-attention.

[0055] Then, a lightweight bottleneck convolutional network is constructed, using 1×1 convolutions. Instead of fully connected layers, a 3×3 depthwise convolution is used to enhance information exchange between adjacent pixels. Extract local information. Finally, use a 1×1 convolution. Replace the second fully connected layer to output a global spectral-spatial feature map. .

[0056] ;

[0057] in, This is the activation function.

[0058] Lightweight bottleneck convolutional networks significantly enhance the model's ability to extract local information, while having lower parameter count and computational complexity.

[0059] In some embodiments of this application, a convolutional map is used to build a decoder to generate a reconstructed image, including: performing channel dimension skipping connections between the deep spectral-spatial feature map output by the encoder and the fine-grained spectral-spatial feature map to obtain a stitched feature; the stitched feature is then generated by two-dimensional convolution through the convolutional map to generate a reconstructed image.

[0060] Furthermore, although the deep latent features extracted by the encoder contain rich high-order semantic information, the fine-grained spectral details of the original hyperspectral image data often suffer irreversible loss as network depth increases and channel dimensions change. To compensate for this information loss and more accurately fit the background data, a cross-layer feature fusion mechanism is introduced in the decoding stage. Its core is to construct a shortcut connecting the low-level physical features with the high-level semantic features. This integrates fine-grained spectral-spatial features... And the deep semantic latent global spectral-spatial feature map Z output by the encoder 2d By performing skip-concatenation along the channel dimension, aggregated features are obtained. :

[0061] ;

[0062] in, ( This refers to the stitching operation along the channel dimension. This design, by introducing primary spectral features as guidance, can significantly reduce the reconstruction error of background pixels.

[0063] Furthermore, to map the aggregated features back to the original data space to generate the reconstructed image and avoid checkerboard effects that may be introduced by cascaded deconvolution decoding, a more direct convolutional mapping strategy is adopted to reduce computational redundancy while maintaining the fineness of the background texture. The mixed features are first weighted and combined by a standard two-dimensional convolutional layer, and then passed through a normalization layer to normalize the feature distribution, generating the reconstructed image. :

[0064] ;

[0065] in, For normalization function, and To reconstruct the weights and biases of the convolutional layer.

[0066] In some embodiments of this application, a two-layer convolutional module is introduced into the pixel-level deep clustering network to perform clustering distribution fitting on the extracted latent spectral-spatial features and output the fitted feature vector. Using this fitted feature vector as a priori, assuming that the spectral-spatial features follow a multivariate Gaussian mixture distribution, the mixing coefficient, mean vector, and covariance matrix of each Gaussian component are solved, and then a Gaussian probability density loss function is constructed. Based on this loss function, the network feature mapping process is regularized, driving background pixels to converge to the high probability density region of the multivariate Gaussian mixture distribution, while increasing the spatial distance between anomalous pixels and the multivariate Gaussian mixture distribution. Pixel-level anomaly classification is performed on each image block. By performing pixel-level anomaly classification around the cluster center of each image block and mapping the latent features back to the image space, clustering modeling based on latent features can be completed.

[0067] Furthermore, to reveal the statistical distribution of the latent features of the modeling background, an independent clustering network is introduced. Used for latent global spectral-spatial feature maps of encoder output Perform cluster distribution fitting to obtain cluster membership degrees. It includes the first convolutional block Conv1 and the second convolutional block Conv2, and its structure diagram is as follows. Figure 5 As shown.

[0068] Furthermore, the first convolutional block Conv1 contains a two-dimensional convolution with a kernel size of 1×1. Its output channel number is the preset number of cluster centers. Four times the original value is then applied, followed by hyperbolic tangent activation (Tanh) activation and normalization operations to perform nonlinear transformation and normalization of the features, preparing for subsequent cluster assignment:

[0069] ;

[0070] in, For normalization function, The hyperbolic tangent activation function is used. and They are respectively The convolution kernel weights and biases, This represents the output features of the first convolutional block.

[0071] Furthermore, the second convolutional block Conv2 contains another two-dimensional convolution with a kernel size of 1×1. The number of output channels is (The value is equal to the preset number of cluster centers). The output is generated by activating a normalized exponential function (Softmax). 3D fitted feature vector :

[0072] ;

[0073] in, To convert any real-valued vector into a normalized exponential function of a probability distribution, and They are respectively The convolution kernel weights and biases, The value of dimensionality is equal to the number of output channels of the second convolutional block. .

[0074] Furthermore, fit the feature vector Each component Indicates the first The pixel belongs to the first The probability of a Gaussian component, where... The pixel ordinal number; Let K be the ordinal number of the Gaussian components, and the sum of the k Gaussian components equals K.

[0075] Furthermore, the fitted feature vectors output by the clustering network are combined... Assuming a potential global spectral-spatial feature map The following can be expressed as a multivariate Gaussian density:

[0076] ;

[0077] in, Let be the probability density function of a multivariate Gaussian mixture distribution. For the first A Gaussian probability density function, and The first The mean vector and covariance matrix of a Gaussian probability density function; Let be the proportion of the k-th Gaussian probability density function in the entire Gaussian mixture model (GMM), i.e., the mixing coefficient of each Gaussian component, and .

[0078] Furthermore, the mixing coefficients of each Gaussian component are solved. Mean vector and covariance matrix Valuation:

[0079] ;

[0080] ;

[0081] ;

[0082] Where N is the total number of pixels in the original hyperspectral image. For the first A pixel-level global spectral-spatial feature map.

[0083] Furthermore, a Gaussian probability density loss function is constructed:

[0084] ;

[0085] in, This represents the transpose of mathematical symbols. Represents the potential global spectral-spatial feature map Dimensions.

[0086] By applying this loss function to the network feature mapping process, the encoder is forced to map background pixels to the high-density region of the GMM, while keeping anomalies away from this distribution.

[0087] In some embodiments of this application, in order to address the problem of scarce labeled samples and to efficiently leverage the performance of the designed backbone feature extraction encoder, a two-stage network optimization strategy consisting of general training and reinforcement training is proposed.

[0088] Furthermore, the general training phase is based on an autoencoder of a hyperspectral anomaly detection network, and the loss is determined by reconstruction error. The process is driven by L2 loss constraints, which first assists in clustering similar pixels by maximizing the preservation of background features. This stage not only enables the transfer of pixel-level classification models to anomaly detection tasks, but also provides highly representative spectral-spatial features for augmented training, thereby effectively improving the separability of the Gaussian mixture model from the background to anomalies.

[0089] ;

[0090] in, To reconstruct the error loss, For 3D domain image blocks, A reconstructed image of a 3D domain image patch.

[0091] Furthermore, the samples for the general training phase can be either highly annotated hyperspectral classification image patches or hyperspectral image patches for anomaly detection, while the training samples for augmented training are only the latter. This phase uses the model weights obtained from general training as initial parameters, effectively improving the model's convergence speed. Simultaneously, this phase minimizes the joint total loss of reconstruction error and Gaussian probability density estimation. This further guides the model to learn the latent statistical distribution of the GMM characterized by the clustering network.

[0092] ;

[0093] in, The Gaussian probability density loss function is used. This is a hyperparameter used to balance the contributions of reconstruction loss and Gaussian probability density loss.

[0094] In some embodiments of this application, the background clustering label of the image block is determined by using the latent feature clustering modeling result output by the pixel-level deep clustering network as a prior constraint, and selecting the category with the optimal probability density as the current pixel clustering label based on the maximum a posteriori probability criterion; the labeling of all pixel categories is completed by traversing the image block pixel by pixel, and finally the background clustering labeling result of the entire image block is output.

[0095] In some embodiments of this application, the process of generating intra-class RX detection scores in the feature space using the RX detection algorithm for each labeled image patch background clustering label is as follows: the image patch is adaptively divided based on the background clustering label of the image patch to obtain several pixel subsets with consistent statistical characteristics; for each pixel subset, the corresponding intra-class mean and intra-class covariance are calculated, and the intra-class RX detection score in the feature space is generated based on the obtained intra-class mean and intra-class covariance.

[0096] To address the challenge of extremely complex backgrounds and a very low proportion of small, anomalous targets in large-size hyperspectral images, the RX decision space is often better suited to the diversity of the background than the separability of the anomalies—that is, the statistical model is entirely dominated by the background, causing the global RX response to weak local anomalies to be overwhelmed by the overall image background. Numerous studies have shown that the response of anomalous targets is significantly higher under highly homogeneous local backgrounds than under global RX. Therefore, this application proposes a method for intra-class anomaly detection in a feature space:

[0097] Based on the Image blocks Clustering results (indicating the first) Fitted feature vectors in image patches Based on the maximum a posteriori probability criterion, the class with the optimal probability density is selected as the first... Clustering labels per pixel :

[0098] ;

[0099] Furthermore, image patches Adaptively divided into K statistically consistent subsets of pixels (K value equal to the number of output channels of the second convolutional block). For each subset of pixels ( The value equals the Gaussian component ordinal number. Extract pixels within the group :

[0100] ;

[0101] in, For the first The first image patch Global spectral-spatial feature map of pixels .

[0102] Set the number of pixels within the class Calculate the mean of this class in the latent space. :

[0103] ;

[0104] Calculate the within-class covariance:

[0105] ;

[0106] Therefore, for those belonging to the pixel subset The For each pixel, its RX score within the latent feature space class is:

[0107] ;

[0108] in, This indicates the transpose of mathematical symbols.

[0109] In some embodiments of this application, while local intra-class RX detection results based on latent space can effectively enhance the response to minor anomalies, their receptive field is small, and the anomaly scores of some feature edges may be inflated. Therefore, this application further introduces an anomaly suppression mechanism based on reconstruction error to construct complementary anomaly detection cues.

[0110] Furthermore, through the image patch-original image aggregation operator The in-class RX scores of the image block-level feature space are respectively and reconstructed image Mapping back to the original image space, global RX anomaly response maps based on in-class RX in the latent space are generated respectively. and reconstructing the full map :

[0111] ;

[0112] ;

[0113] in, Implicit representations include operations such as averaging and fusion of overlapping regions across all image patches in spatial location. This represents the total number of image patches. This is the sequence number of a specific image patch.

[0114] In some embodiments of this application, the decision fusion process of the global RX anomaly response map and the reconstructed full map is as follows: calculate the reconstruction error between the original hyperspectral image and the reconstructed full map to generate a weight matrix, use the weight matrix to perform adaptive weighted modulation on the global RX anomaly response map, and finally output the hyperspectral image anomaly detection result map.

[0115] Furthermore, the hyperspectral feature map S after band selection is used to reconstruct the full map. Construct a second anomaly metric based on the reconstruction error between them. :

[0116] ;

[0117] Subsequently Normalization is performed, and the weight matrix is ​​generated using the normalization exponential function (Softmax):

[0118] ;

[0119] in, Indicates the first The second anomaly measure for each pixel. Indicates from the first From one pixel to the total number of pixels in the hyperspectral raw image The sum of abnormality measures.

[0120] Finally, the global anomaly response map is analyzed using a weight matrix guided by the reconstruction error. Modulation is performed to suppress false alarm responses in the background region, resulting in the final anomaly detection score, and finally outputting the hyperspectral image anomaly detection result map.

[0121] ;

[0122] To verify the technical effectiveness of the technical solution provided in the embodiments of this application, the following experiment was designed: The hyperspectral raw image data used in the experiment came from the Advanced Visible-Shortwave Infrared Hyperspectral Imager (GF5-AHSI) carried by the Gaofen-5 satellite of China. To verify the performance of the algorithm under extremely harsh conditions, such as... Figure 6As shown, this application selected four raw shortwave infrared hyperspectral images with different background types as experimental data. Each image is 512×512 pixels in size and contains 180 spectral bands. The first HSI image, denoted Gf5_1, has a predominantly forested and cloudy background. The second HSI image, denoted Gf5_2, has a predominantly ocean background with thin clouds. The third image, denoted Gf5_3, has a predominantly ocean and near-shore land background. The fourth HSI image, denoted Gf5_4, has a predominantly cloudless mountainous and grassland background. It is noteworthy that the anomalies in Gf5_1 through Gf5_4 are all pixel-level micro-targets, accounting for only 0.002% of the total pixels.

[0123] To verify the performance of the anomaly detection method proposed in this application, it is comprehensively compared with various types of hyperspectral anomaly detection methods in the prior art, including: the classic representative of traditional statistical methods RX, the classic method CBAD based on clustering models, the standard method TLRSR based on representation learning, two classic methods (DirectNet and PUNNet) based on replacing blind spot frames with local window center pixels, GT-HAD based on local patch reconstruction and aggregation to generate whole image residuals, and two classic methods (SGLNet and BSRegNet) based on whole image background residual reconstruction models.

[0124] Performance evaluation employed color anomaly maps, receiver operating characteristic (ROC) curves, and the area under the ROC curve (AUC). PD-PF The area under the curve (AUC) of combined anomaly detectability versus background suppressability TD-BS The metrics include box plots and other indicators. Among these, the color anomaly map visually reflects the method's detection performance in terms of anomaly response and background suppression; the ROC curve effectively describes the relationship between the probability of detection (PD) and the false alarm rate (PF)—the better the detector's performance, the closer its ROC curve is to the upper left of the image, and the higher its corresponding AUC. PD-PF The closer the value is to 1, the better; TD-BS By using the target detectability AUC at the same threshold τ (PD,τ) Background false alarm AUC (PF,τ) Subtract to evaluate the overall effectiveness of the detector in suppressing background noise (Formula: Box plots are used to reflect the separability between the background and the anomaly.

[0125] refer to Figure 6A comparison of the anomaly detection color anomaly maps provided in this application with various existing hyperspectral anomaly detection methods on the GF5-AHSI experimental dataset reveals that: on the large-scale GF5 dataset, the three non-deep learning methods, RX, CBAD, and TLRSR, have weak background suppression capabilities and can even amplify noise, interfering with anomaly detection in complex backgrounds, especially in cloud-containing Gf5_1 and Gf5_2 scenes, where their detection accuracy is lower than that of deep learning methods. DirectNet, PUNNet, and SGLNet methods can significantly suppress background, but they cannot effectively highlight anomalous targets and may mistakenly treat anomalies as background weakening. GT-HAD and BSRegNet methods can eliminate background interference and enhance anomaly saliency in relatively uniform background scenes, but in strongly changing scenes, local complex backgrounds are easily reconstructed and difficult to distinguish from anomalies. The method proposed in this application significantly amplifies the discriminative difference between anomalous targets and local backgrounds, making anomaly detection response most prominent in various scenes.

[0126] refer to Figure 7 The ROC curves of the method provided in this application and various existing hyperspectral anomaly detection methods on the GF5-AHSI experimental dataset are compared, where (a) corresponds to the Gf5_1 dataset, (b) corresponds to the Gf5_2 dataset, (c) corresponds to the Gf5_3 dataset, and (d) corresponds to the Gf5_4 dataset; combined with the AUC of the method proposed in this application and various existing hyperspectral anomaly detection methods on the GF5-AHSI experimental dataset in Table 1. PD-PF With AUC TD-BS Performance comparison shows that the ROC curves of this method in all four differentiated scenarios are close to the upper left corner of the coordinate axis, and exhibit the steepest upward trend in the initial stage, verifying the strong adaptability of the background clustering constraint and latent feature separation mechanism to complex surface environments. Other algorithms' AUC... PD-PF While other methods struggle to maintain a detection accuracy above 0.95, our method consistently achieves over 0.995 on all datasets, demonstrating exceptionally high accuracy and robustness. Furthermore, our method achieves a high AUC (Average Value of 1) in conjunction with background suppression and target detection. TD-BS Furthermore, the performance of the method in this application can basically maintain the optimal level, solving the industry pain point of "abnormal loss caused by excessive background suppression".

[0127] Table 1. Comparison of anomaly detection performance of the proposed method with various existing hyperspectral anomaly detection methods on the GF5-AHSI experimental dataset.

[0128] The best and second-best performance for each dataset is shown in bold and underlined, with ↑ indicating that a larger value is better.

[0129] refer to Figure 8 The box-separable plots of anomaly detection using the method provided in this application and various existing hyperspectral anomaly detection methods on the GF5-AHSI experimental dataset are compared. (a) corresponds to the Gf5_1 dataset, (b) to the Gf5_2 dataset, (c) to the Gf5_3 dataset, and (d) to the Gf5_4 dataset. The horizontal axis is labeled: 1: RX, 2: CBAD, 3: TLRSR, 4: DirectNet, 5: PUNNet, 6: GTHAD, 7: SGLNet, 8: BSRegNet, 9: this method. It can be seen that on all test datasets, the anomaly boxes of the RX, CBAD, DirectNet, PUNNet, and BSRegNet methods show significant overlap with the background boxes, while the gaps between the boxes of other comparison methods are also very small. Although the TLRSR algorithm has the highest anomaly boxes on multiple datasets, its anomaly boxes are also significantly taller, easily generating false alarms. The average anomaly box value of the method proposed in this application ranks among the top two of all methods. Furthermore, BCLFS can control the background pixel values ​​within a very low range, ensuring the largest gap between the background bounding box and the anomaly on the six datasets, thus easily extracting the anomaly from the suppressed background. In summary, the method proposed in this application achieves a satisfactory balance between anomaly detectability and background suppression, demonstrating excellent ability to separate the background from the anomaly.

[0130] In summary, the method proposed in this application achieves comprehensive superiority in both detection accuracy and robustness. On the entire GF5-AHSI dataset, covering complex scenarios with strong interference such as cloudy forests and coastal zones, the AUC is [missing value]. PD-PF All metrics exceeded 0.995, effectively overcoming the technical challenge of unstable anomaly detection accuracy in complex terrain environments. Simultaneously, this method achieves a balanced adaptation between background suppression and target enhancement capabilities, relying on an image block-level intra-class detection mechanism, with AUC... TD-BS The metrics maintained optimal levels, effectively addressing the industry pain point of "abnormal target loss caused by excessive background suppression." Furthermore, the model's adaptability to complex scenarios was significantly improved, exhibiting ideal performance in diverse scenarios, fully validating the superior adaptability of the background clustering constraint and latent feature separation mechanism proposed in this application to large-scale complex terrain scenarios.

[0131] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0132] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A hyperspectral anomaly detection method based on background clustering constraint and potential feature separation, characterized in that, include: Step 1: Perform band selection processing on the original hyperspectral image data and generate fixed-size image blocks through spatial partitioning; Step 2: Construct a hyperspectral anomaly detection network and build an encoder using a lightweight spectral-spatial feature extractor to mine global semantic information and deep spectral-spatial features of image patches; A skip connection mechanism is introduced to transmit the stitched spectral-spatial features to the decoding end, and a convolutional mapper is used to build a decoder to generate the reconstructed image; Step 3: A pixel-level deep clustering network is introduced between the encoder and decoder. A Gaussian mixture model is applied to the latent spectral-spatial features extracted by the pixel-level deep clustering network to perform latent constraints. The latent feature clustering model is completed based on the Gaussian probability density loss function, thereby labeling the background clustering labels of the image blocks. For the labeled background clustering labels of the image blocks, the RX detection algorithm is applied to generate the intra-class RX detection scores in the feature space. Step 4: The intra-class RX detection scores in the feature space are mapped back to the original hyperspectral image using the aggregation operator to form a global RX anomaly response map and a reconstructed full image. Step 5: Perform decision fusion on the global RX anomaly response map and the reconstructed full map to form the final hyperspectral image anomaly detection result map.

2. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The band selection process involves calculating the correlation coefficients of adjacent bands in the original hyperspectral image data to determine the correlation distribution of the original hyperspectral image data. Based on the minimum value of the correlation coefficients of adjacent bands and a set threshold, the original hyperspectral image data is adaptively divided into several data subspaces. All band indices in each data subspace are calculated, and a preset number of bands with the largest band indices are selected as representative bands in each data subspace in the same proportion. The representative bands selected in each data subspace are arranged in the initial band order to form a hyperspectral feature map after band selection.

3. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 2, characterized in that: The spatial partitioning uses a window of a certain size and slides on the hyperspectral feature map after band selection processing with a certain sliding step size to extract overlapping three-dimensional neighborhood image patches and generate a dataset composed of several three-dimensional neighborhood image patches.

4. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The process of constructing an encoder using a lightweight spectral-spatial feature extractor to mine global semantic information and deep spectral-spatial features of image patches is as follows: A 3D convolutional module is used to extract joint features of the spectral and spatial dimensions of the image patches, outputting a local spectral-spatial feature map; a 2D convolutional module is introduced to extract features from the local spectral-spatial feature map output by the 3D convolutional module again, generating a feature map after spatial dimension reconstruction and enhancement; based on the Transformer architecture, the feature map after spatial dimension reconstruction and enhancement is normalized and input into a multi-head self-attention mechanism to establish long-distance dependencies, generating a global spectral-spatial feature map; a lightweight bottleneck convolutional network is constructed to enhance the local spatial modeling capability of the global spectral-spatial feature map.

5. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: A decoder is built using a convolutional map to generate a reconstructed image, including: performing channel dimension skipping connections between the deep spectral-spatial feature map output by the encoder and the fine-grained spectral-spatial feature map to obtain the stitched spectral-spatial feature map; and then performing two-dimensional convolution on the stitched spectral-spatial feature map to generate a reconstructed image.

6. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The pixel-level deep clustering network introduces a two-layer convolutional module to perform clustering distribution fitting on the extracted latent spectral-spatial features and output a fitted feature vector. Using this fitted feature vector as a priori, assuming that the spectral-spatial features follow a multivariate Gaussian mixture distribution, the mixing coefficients, mean vectors, and covariance matrices of each Gaussian component are solved, and a Gaussian probability density loss function is constructed. Based on this loss function, the network feature mapping process is regularized, driving background pixels to converge to the high probability density region of the multivariate Gaussian mixture distribution, while increasing the spatial distance between anomalous pixels and the multivariate Gaussian mixture distribution. By performing pixel-level anomaly classification around the cluster center of each image patch and mapping the latent features back to the image space, the latent feature clustering modeling can be completed.

7. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The background clustering labels of image blocks are calibrated by using the latent feature clustering modeling results output by the pixel-level deep clustering network as a prior constraint, and selecting the class with the optimal probability density as the current pixel clustering label based on the maximum a posteriori probability criterion. The labeling of all pixel class labels is completed by traversing the image block pixel by pixel, and finally the calibration results of the background clustering labels of the entire image block are output.

8. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The process of generating intra-class RX detection scores in the feature space using the RX detection algorithm is as follows, based on the background clustering labels of each labeled image patch: the image patch is adaptively divided based on the background clustering labels of the image patch to obtain several pixel subsets with consistent statistical characteristics; for each pixel subset, the corresponding intra-class mean and intra-class covariance are calculated, and the intra-class RX detection scores in the feature space are generated based on the obtained intra-class mean and intra-class covariance.

9. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The decision fusion process of the global RX anomaly response map and the reconstructed full map is as follows: calculate the reconstruction error between the original hyperspectral image and the reconstructed full map to generate a weight matrix, use the weight matrix to perform adaptive weighted modulation on the global RX anomaly response map, and finally output the hyperspectral image anomaly detection result map.

10. The hyperspectral anomaly detection method based on latent feature separation under background clustering constraints according to claim 1, characterized in that: The hyperspectral anomaly detection network is optimized using a two-stage training strategy, including: General training phase: Based on the autoencoder structure of the hyperspectral anomaly detection network, the reconstruction process is constrained by the L2 loss function. By maximizing the preservation of background features to assist in clustering similar pixels, the pixel-level classification model is transferred to the anomaly detection task, and spectral-spatial features with strong representational power are obtained. Enhanced training phase: Using the model weights output from the general training phase as initial parameters, the hyperspectral image patch to be detected is used as training sample. By minimizing the joint loss function of reconstruction error and Gaussian probability density estimation, the model is guided to learn the latent statistical distribution of Gaussian mixture model characterized by pixel-level deep clustering network. The samples in the general training phase are selected from hyperspectral classification image patches with fine annotations or hyperspectral image patches to be detected for anomalies. The joint loss function consists of reconstruction error loss and Gaussian probability density loss function.

Images