Hyperspectral anomaly detection method and device based on encoding and independent target subspace projection

Abstract

The present invention relates to the field of image processing, and discloses a hyperspectral anomaly detection method and device based on encoding and independent target subspace projection. The method comprises the following steps: (1) determining the number of neurons in the hidden layer of an autoencoder and the number of independent targets; (2) training the original data using the autoencoder to obtain an encoding weight matrix and an encoding space projection matrix; (3) projecting the original data onto an orthogonal subspace of the encoding space to obtain orthogonal components of low-rank background components; (4) sphering the orthogonal components of the low-rank background components, and using them to search for independent targets for constructing an independent target subspace; (6) projecting the sphered orthogonal components onto the independent target subspace to obtain independent target components; (7) using sparse cardinality to constrain the independent target components to obtain anomaly components, combining them with a distance operator to construct a detection operator to achieve anomaly detection. The present invention utilizes the autoencoder and independent target space projection to effectively suppress background and enhance targets.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a hyperspectral anomaly detection method and device based on encoding and independent target subspace projection. Background Art

[0002] Hyperspectral remote sensing images offer outstanding features such as unified image and spectrum and high spectral resolution. Their rich spectral information can identify surface materials with diagnostic spectral characteristics and detect materials that are undetectable in high-spatial-resolution images like panchromatic and multispectral images. This makes them superior to other remote sensing imagery in target detection and recognition. Hyperspectral anomaly detection (HAD) is the process of detecting potentially suspicious materials in hyperspectral imagery (HSI) without prior spectral information about the target. It is widely used in military surveillance, precision agriculture, and rare mineral exploration.

[0003] Although there is no prior information about anomalous targets, they possess the following characteristics: Based on statistical characteristics, they are assumed to occupy a small number of pixels, have a low probability of occurrence, and are unpredictable; and based on spectral differences, they are assumed to differ from their surrounding pixels. Over the past few decades, much research has been conducted based on these two prerequisites. Unlike object detection, which relies on prior knowledge of the target, HAD relies more on the suppression effect of the background (BKG) and the contrast between the BKG and the anomaly.

[0004] To better achieve background suppression and target enhancement, a method based on low-rank and sparse representation (LRaSR) decomposes the raw data into background + target + noise components and utilizes an optimized parameter solution strategy to obtain the target component while addressing the background and noise issues. For LRaSR-based methods, the low-rank background component is essentially obtained by finding an optimal low-rank space to represent the back-knotted gated gate (BKG). There are two key issues to address in finding this optimal low-rank space: how to determine the cardinality m used to represent the rank of the BKG; and how to obtain the m spatial basis vectors. Most methods rely on certain prior distribution assumptions, such as GoDec, which is based on singular / eigenanalysis techniques and assumes that the BKG follows a Gaussian distribution. Furthermore, the low-rank constraints in parameter optimization methods are converted to the calculation of the matrix F-norm, which also uses singular analysis concepts. However, the background component of real hyperspectral data does not simply satisfy this distribution assumption. Therefore, a low-rank BKG space can be obtained by extracting or finding a valid representation of the raw data.

[0005] In recent years, with the rapid development of deep learning applications in various fields, hyperspectral image processing techniques based on deep learning have also garnered widespread attention. Among them, the autoencoder (AE), as a type of unsupervised neural network, can capture abstract features to represent the raw HSI and reconstruct it without prior distribution assumptions, making it suitable for unsupervised HAD. Currently, common methods utilize autoencoder latent features, reconstruction errors, or constrained reconstruction errors for anomaly detection. However, while these methods avoid the assumption of a prior distribution of background, they still face challenges in determining the cardinality of background representation and noise interference. Unlike these methods based on reconstruction errors or latent features, this approach, driven by both data and AE models, leverages the low pixel count and probability of anomaly targets. By constraining the number of AE latent neurons, the network is biased towards reconstructing the background, thereby obtaining a background encoding space for background component construction. Unsupervised target detection within the sphericalized background orthogonal components achieves background suppression and Gaussian noise removal while simultaneously obtaining an independent target space, contributing to improved background suppression and target enhancement. In order to further reduce the impact of non-Gaussian noise, sparse cardinality constraints are imposed on independent target components to obtain abnormal components, and the final detection results are obtained by combining the distance operator. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the present invention aims to provide a hyperspectral anomaly detection method and device based on encoding and independent target subspace projection.

[0007] Taking full advantage of the characteristics of hyperspectral data and the advantages of autoencoders in data representation, by constraining the number of latent neurons of the autoencoder, the autoencoder network is driven to tend to reconstruct the background rather than abnormal targets and noise. That is, the present invention combines model-driven and data-driven, uses a constrained network architecture to obtain a data coding space, and projects the original data into a low-rank coding space to retain background information. In the orthogonal subspace of the low-rank background space, unsupervised target detection of the sphericalized data components can suppress the background, remove Gaussian noise and obtain independent target components. Combined with sparse cardinality constraints, the influence of non-Gaussian noise is further reduced to obtain abnormal components and enhance the target. Finally, the distance operator and abnormal components are combined to achieve anomaly detection.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The present invention first provides a hyperspectral anomaly detection method based on encoding and independent target subspace projection, which is characterized by comprising the following steps:

[0010] Step 1): Calculate the background rank m and the number of independent targets j of the original hyperspectral data X. The background rank m is used to determine the number of neurons in the hidden layer of the autoencoder in step 2), and the independent target number j is used to determine the sparse cardinality in step 5);

[0011] Step 2): Train the autoencoder with background rank constraint to obtain the encoder weight matrix and the encoding space projection matrix

[0012] Step 3): Using the encoding space projection matrix Project the original hyperspectral data X into the encoding space Orthogonal subspace of Get the low-rank background component L m The orthogonal components

[0013] Step 4): In order to obtain independent targets for independent target space construction, the orthogonal components of the low-rank background components are sphericalized, and the sphericalized orthogonal components are used for independent target search, and the searched independent target set is used for the independent target subspace structure;

[0014] Step 5): The low-rank background component L after spherization m The orthogonal components Projection to independent target subspace Obtain independent target components and use sparse cardinality constraints to obtain abnormal components S j , so far the original hyperspectral data is decomposed into the orthogonal sum of low-rank background component, abnormal component and noise component;

[0015] Step 6): Utilize the abnormal component S j Combined with Mahalanobis distance to construct detection operator Realize the detection of abnormal targets.

[0016] The present invention also provides a hyperspectral anomaly detection device based on encoding and independent target subspace projection for implementing the above method, comprising:

[0017] A background rank number and independent target number determination module is used to determine the background rank number and the number of independent targets in the original data. The background rank number is used to determine the number of latent neurons in the process of constructing the autoencoder in the low-rank background component generation module, and the independent target number is used to determine the sparse cardinality in the abnormal component generation module;

[0018] A low-rank background component generation module is used to construct and train an autoencoder, use the encoding matrix of the autoencoder to construct an encoding space projection matrix, and use the projection matrix to project the original data into the encoding space to obtain a low-rank background component;

[0019] A low-rank background orthogonal component generation module is used to project the original hyperspectral data into the orthogonal subspace of the low-rank encoding space of the original data to obtain the orthogonal components of the low-rank background components, i.e., the low-rank background orthogonal components;

[0020] The data spheroidization module is used to spheroidize the input low-rank background orthogonal components to obtain spheroidized data;

[0021] An independent target component generation module is used to project the sphericalized low-rank background orthogonal component to the independent target component subspace of the sphericalized low-rank background orthogonal component to obtain the independent target component;

[0022] An abnormal component generation module, used for applying sparse cardinality constraints to independent target components to obtain abnormal components;

[0023] The detection operator module is used to combine abnormal components with distance operators to construct an anomaly detection operator;

[0024] The anomaly detection result output module is used to output the anomaly detection result graph.

[0025] The beneficial effects of the present invention are:

[0026] 1) The present invention constrains the number of latent neurons in the AE network by adaptively determining the background rank number, and uses the encoding matrix obtained after training to construct an encoding space. The original data is projected into the encoding space to obtain low-rank background components to characterize the background. Adaptively determining the number of latent neurons in the AE network can drive the network to tend to reconstruct the background rather than abnormal targets and noise. The proposed BKG representation method combines the advantages of data-driven and network structure constraints, which is different from other AE-based AD methods that utilize the reconstruction error or potential features of the AE framework.

[0027] 2) The original data is projected onto an orthogonal subspace of the low-rank background space. The resulting components are then spherified and extracted to construct an independent target space. This operation not only suppresses the background but also eliminates the influence of Gaussian noise, obtaining independent target components that contain anomalous targets. To further reduce the influence of non-Gaussian noise, a sparse cardinality constraint is imposed on the independent target components, which improves target enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flowchart of the basic steps of an embodiment of the hyperspectral anomaly detection method of the present invention;

[0029] Figure 2 Schematic diagram of the structure of the hyperspectral anomaly detection device of the present invention;

[0030] Figure 3Hypersion hyperspectral images are used for experiments;

[0031] Figure 4 A comparison chart of anomaly detection results of Hypersion hyperspectral images using different detection methods. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to specific embodiments and drawings, and specific embodiments are described to simplify the present invention. However, it should be understood that the present invention is not limited to the embodiments described, and various modifications of the present invention are possible without departing from the basic principles, and these equivalent forms also fall within the scope defined by the claims appended hereto.

[0033] First, some basic variables related to the present invention are introduced.

[0034] Let the original data be X, the background rank be m, the number of independent targets be j, and the sparse cardinality be k. The encoding matrix of the autoencoder is The projection matrix of the encoding space is The m-rank coding space is The m-rank background component is divided into L m , the orthogonal subspace projection matrix of the m-rank coding space is The orthogonal components of the background components are

[0035] like Figure 1 FIG. 1 is a flowchart of the basic steps of the hyperspectral anomaly detection method based on encoding and independent target subspace projection in this embodiment of the present invention, which mainly includes:

[0036] Step 1: Determine the background rank and independent target number of the original hyperspectral data: Calculate the number of virtual endmembers p, background rank m, and independent target j of the original data X;

[0037] For the original data X, the NWHFC algorithm is used to determine the number of virtual endmembers p. MX-SVD is used to further divide p to obtain the background rank m and the number of independent targets j.

[0038] Step 2: Use the autoencoder to extract the latent features of the original data X, where the autoencoder is a neural network consisting of only one latent layer (containing m neurons), with the input being the original data X and the output being the reconstructed data X′. Project X into the low-rank encoding space of the original data to obtain the m-rank background component; use the encoding matrix of the original data X Construct encoding space projection matrix And project the original data X into the low-rank encoding space to obtain the low-rank background component

[0039] Step 3: Project X to the orthogonal subspace of the encoding space to obtain the orthogonal components of the low-rank background components; project the original data X to the encoding space Orthogonal subspace of The projection matrix Orthogonal components of background components

[0040] Step 4: Spherization And use the unsupervised target extraction method ATGP to obtain j independent target vectors Construct an independent target vector matrix T j =[t1,t2…t j ], using the independent target vector matrix T j Constructing the projection matrix Orthogonal components after spherization Projection to independent target space Obtain independent target components

[0041] Step 5: Using the number of independent component components j, impose a sparse cardinality constraint on the independent target group, that is, retain the largest k elements in the independent target component, k = j * N, k is the sparse cardinality, N is the number of known hyperspectral image pixels, and the calculation formula for the obtained abnormal components is:

[0042] The abnormal components obtained P Ω (·) represents the projection matrix that projects the matrix in the brackets onto the element set Ω, where Ω represents the non-zero set of the first k largest elements in the matrix.

[0043] Step 6: Using the obtained abnormal component S j Combined with Mahalanobis distance to construct detection operator The calculation formula of the anomaly detection operator is Substituting the pixel to be detected into the operator formula can obtain the pixel anomaly degree, and traversing all pixels can complete the detection of abnormal targets.

[0044] Corresponding to the aforementioned embodiment of a hyperspectral anomaly detection method based on encoding and independent target subspace projection, the present invention also provides an embodiment of a hyperspectral anomaly detection device based on encoding and independent target subspace projection.

[0045] Figure 2 FIG. 1 is a block diagram of a hyperspectral anomaly detection method based on encoding and independent target subspace projection according to an exemplary embodiment. Figure 2 As shown, the device includes:

[0046] A background rank number and independent target number determination module is used to determine the background rank number and the number of independent targets in the original data. The background rank number is used to determine the number of latent neurons in the process of constructing the autoencoder in the low-rank background component generation module, and the independent target number is used to determine the sparse cardinality in the abnormal component generation module;

[0047] A low-rank background component generation module is used to construct and train an autoencoder, use the encoding matrix of the autoencoder to construct an encoding space projection matrix, and use the projection matrix to project the original data into the encoding space to obtain a low-rank background component;

[0048] A low-rank background orthogonal component generation module is used to project the original hyperspectral data into the orthogonal subspace of the low-rank encoding space of the original data to obtain the orthogonal components of the low-rank background components, i.e., the low-rank background orthogonal components;

[0049] The data spheroidization module is used to spheroidize the input low-rank background orthogonal components to obtain spheroidized data;

[0050] An independent target component generation module is used to project the sphericalized low-rank background orthogonal component to the independent target component subspace of the sphericalized low-rank background orthogonal component to obtain the independent target component;

[0051] An abnormal component generation module, used for applying sparse cardinality constraints to independent target components to obtain abnormal components;

[0052] The detection operator module is used to combine abnormal components with distance operators to construct an anomaly detection operator;

[0053] The anomaly detection result output module is used to output the anomaly detection result graph.

[0054] In a specific implementation of the present invention, an application expansion module is further included, which is used to use low-rank background components and abnormal components to perform hyperspectral image classification, unmixing or target detection.

[0055] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0056] As for the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is only schematic. For example, the low-rank background component module can be divided into a logical function. There may be other division methods in actual implementation. For example, multiple modules can be combined or integrated into another unit. Another point is that the connection between the modules shown or discussed can be a communication connection through some interfaces, which can be electrical or other forms. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present application scheme. Those of ordinary skill in the art can understand and implement it without paying creative labor. The following takes real hyperspectral images as an example to illustrate a specific implementation method to reflect the technical effect of the present invention. The specific steps in the embodiment will not be repeated.

[0057] Example

[0058] Next, we use Hypersion hyperspectral images as the research object to carry out the verification of anomaly detection algorithm. In order to compare the anomaly detection results from the two perspectives of intuitive visual and quantitative evaluation, we use the anomaly detection result graph and 3D-ROC and its derived evaluation indicators including AUC. TD , AUC BS , AUC SNPR , AUC TDBS and AUC ODP Etc., the detection results are evaluated from the perspectives of target detection ability, background suppression ability, and detector effectiveness. 3D-ROC includes AUC (D,F) , AUC (D,t) , AUC (F,t) , respectively characterizing the detector effectiveness, target detection capability, and background suppression capability. The derived evaluation indicators are also divided according to the corresponding evaluation angles, as follows:

[0059] AUC TD represents the joint detection capability of the detector,

[0060] 0≤AUC TD =AUC (D,F) +AUC (D,τ) )≤2

[0061] AUC BS represents the joint background suppression capability of the detector,

[0062] -1≤AUC BS =AUC (D,F) -AUC (F,τ) )≤1

[0063] AUC TDBSrepresents the detector’s comprehensive ability of target detection and background suppression,

[0064] -1≤AUC TDBS =AUC (D,τ) -AUC (F,τ) )≤1

[0065] AUC SNPR represents the signal-to-noise ratio of the detector,

[0066]

[0067] AUC ODP Represents the total detection probability of the detector

[0068] -1≤AUC ODP =AUC (D,F) +AUC (D,τ) -AUC (F,τ) ≤2

[0069] In general, the evaluation indicators are divided as follows:

[0070] (a) Target detection capability (TD): AUC (D,t) ,AUC TD

[0071] (b) Background suppression ability (BS): AUC (F,t) ,AUC BS ,AUC SNPR

[0072] (c) Detector effectiveness: AUC (D,F) ,AUC TDBS ,AUC ODP

[0073] Real hyperspectral images, that is, Hypersion hyperspectral images such as Figure 3 As shown in Figure 3(a), a false color image of a small-width experimental data clip is shown, and Figure 3(b) is a label image of a true anomaly target. Table 1 lists the Hypersion data parameters. Table 2 shows the quantitative indicators of anomaly detection results using EaIT-SP, RX-AD, CRD-DW-STO, OSPDS-AD, LSDM-MoG, PTA, and RGAE on Hypersion hyperspectral images.

[0074] Table 1 Hyperspectral image parameters

[0075] data Sub-region size Number of bands Number of target points Hypersion high spectrum image 100×100 224 / 145 (removes noise and water vapor absorption bands) 32

[0076] Table 2 Quantitative indicators of anomaly detection results of Hypersion hyperspectral images using EaIT-SP, RX-AD, CRD-DW-STO, OSPDS-AD, LSDM-MoG, PTA and RGAE

[0077]

[0078] Figure 4 The comparison chart of anomaly detection results of Hypersion urban hyperspectral images using EaIT-SP, RX-AD, CRD-DW-STO, OSPDS-AD, LSDM-MoG, PTA and RGAE methods is shown, combined with the quantitative detection results of the eight evaluation indicators shown in Table 2. It can be seen that if the evaluation is conducted from the perspective of detector effectiveness, the EaIT-SP method proposed in this invention has a better detection accuracy than the RX-AD method in AUC. (D,F) and AUC BS These two indicators are slightly lower than OSPDS-AD, but overall, both background suppression and target enhancement are better than the comparison algorithm. In summary, the superiority of the anomaly detection operator proposed in this invention in target enhancement and background suppression can be reflected at both the visual effect and quantitative indicators.

[0079] The above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and many variations are possible. All variations that can be directly derived or imagined by a person skilled in the art from the disclosure of the present invention should be considered to be within the scope of protection of the present invention.

Claims

1. Hyperspectral anomaly detection method based on encoding and independent target subspace projection, characterized by: The steps include: Step 1): Calculate the background rank m and the number of independent targets j of the original hyperspectral data X. The background rank m is used to determine the number of neurons in the hidden layer of the autoencoder in step 2), and the independent target number j is used to determine the sparse cardinality in step 5); Step 2): Train the autoencoder with background rank constraint to obtain the encoding matrix of the encoder and the encoding space projection matrix Step 3): Using the encoding space projection matrix Project the original hyperspectral data X into the encoding space Orthogonal subspace of Get the low-rank background component L m The orthogonal components Step 4): In order to obtain independent targets for independent target space construction, the orthogonal components of the low-rank background components are sphericalized, and the sphericalized orthogonal components are used for independent target search, and the searched independent target set is used for the independent target subspace structure; Step 5): The low-rank background component L after spherization m The orthogonal components Projection to independent target subspace Obtain independent target components and use sparse cardinality constraints to obtain abnormal components S j , so far the original hyperspectral data is decomposed into the orthogonal sum of low-rank background component, abnormal component and noise component; Step 6): Utilize the abnormal component S j Combined with Mahalanobis distance to construct detection operator Realize the detection of abnormal targets.

2. The hyperspectral anomaly detection method based on encoding and independent target subspace projection according to claim 1, characterized in that: In step 1), NWHFC combined with Min-Max SVD is used to obtain the background rank m and independent target number j of the hyperspectral data X.

3. The hyperspectral anomaly detection method based on encoding and independent target subspace projection according to claim 1, characterized in that: The step 2) is specifically as follows: The original hyperspectral data X is trained using an autoencoder, where the autoencoder contains a single hidden layer and the hidden layer contains m neurons. The loss function is the root mean square error between the original hyperspectral data and the reconstructed data. The network training uses the quantized conjugate gradient descent method to obtain the encoder matrix and the encoding space projection matrix The calculation formula is:

4. The hyperspectral anomaly detection method based on encoding and independent target subspace projection according to claim 1, characterized in that: The step 3) is specifically as follows: Encoding matrix of raw hyperspectral data obtained using autoencoder Construct the projection matrix with the identity matrix I The calculation formula is: in, Represents the encoding matrix of the autoencoder, and the superscript T represents the transpose; Project the original hyperspectral data X into the low-rank background space Get the low-rank background component L m Project the original hyperspectral data X into the low-rank background space Orthogonal subspace of Get the orthogonal components of the background components The calculation formula is:

5. The hyperspectral anomaly detection method based on encoding and independent target subspace projection according to claim 1, characterized in that: The step 4) is specifically as follows: right Spherization And use the unsupervised target extraction method ATGP to obtain j independent target vectors Construct an independent target vector matrix T j =[t1,t2…t j ]; Using the independent target vector matrix T j Constructing the projection matrix The calculation formula is: Orthogonal components after spherization Projection to independent target subspace Obtain independent target components The calculation formula is:

6. The hyperspectral anomaly detection method based on encoding and independent target subspace projection according to claim 1, characterized in that: The step 5) is specifically as follows: Using the number of independent targets j, a sparse cardinality constraint is imposed on the independent target components, that is, the largest k elements in the independent target components are retained, k = j * N, k is the sparse cardinality, N is the number of known hyperspectral image pixels, and the calculation formula for the obtained abnormal components is: Among them, P Ω (·) represents the projection matrix that projects the matrix in the brackets onto the element set Ω, where Ω represents the non-zero set of the first k largest elements in the matrix, Among them, g r,c,l is the grayscale of the pixel at coordinate position (r,c,l); The original hyperspectral data X is decomposed into the orthogonal sum of three components, namely X=L m +S j +N EaIT-SP (9) Among them, L m is the low-rank background component, S j is an abnormal component, N EaIT-SP The noise component.

7. The hyperspectral anomaly detection method based on encoding and independent target subspace projection according to claim 1, characterized in that: The step 6) is specifically as follows: Using the obtained abnormal component S j Combined with Mahalanobis distance to construct detection operator Calculate each abnormal component S j Pixel vector The abnormal degree, traversing the abnormal component S j All pixels of are used to detect abnormal targets. The calculation formula of the anomaly detection operator is as follows: in, Indicates abnormal component S j The pixel mean vector of Indicates abnormal component S j The inverse of the covariance matrix of ; the superscript T indicates the transpose.

8. A hyperspectral anomaly detection device based on encoding and independent target subspace projection that implements the method of claim 1, characterized in that: include: A background rank number and independent target number determination module is used to determine the background rank number and the number of independent targets in the original data. The background rank number is used to determine the number of latent neurons in the process of constructing the autoencoder in the low-rank background component generation module, and the independent target number is used to determine the sparse cardinality in the abnormal component generation module; A low-rank background component generation module is used to construct and train an autoencoder, use the encoding matrix of the autoencoder to construct an encoding space projection matrix, and use the projection matrix to project the original data into the encoding space to obtain a low-rank background component; A low-rank background orthogonal component generation module is used to project the original hyperspectral data into the orthogonal subspace of the low-rank encoding space of the original data to obtain the orthogonal components of the low-rank background components, i.e., the low-rank background orthogonal components; The data spheroidization module is used to spheroidize the input low-rank background orthogonal components to obtain spheroidized data; An independent target component generation module is used to project the sphericalized low-rank background orthogonal component to the independent target component subspace of the sphericalized low-rank background orthogonal component to obtain the independent target component; An abnormal component generation module, used for applying sparse cardinality constraints to independent target components to obtain abnormal components; The detection operator module is used to combine abnormal components with distance operators to construct an anomaly detection operator; The anomaly detection result output module is used to output the anomaly detection result graph.

9. The hyperspectral anomaly detection device based on encoding and independent target subspace projection according to claim 8, characterized in that: Also included is an application extension module for using low-rank background components and anomaly components to perform hyperspectral image classification, unmixing, or target detection.

Images