Hyperspectral anomaly detection method based on spatial-spectral synergy guided joint dictionary construction

Abstract

The present invention proposes a hyperspectral anomaly detection method that uses spatial-spectral collaborative guidance to construct a joint dictionary, which is used to solve the problem of insufficient mining of background spectrum and abnormal spectrum information and noise interference when constructing the dictionary, resulting in poor detection performance. It includes: 1) using a spatial detector and a spectral detector to detect the input hyperspectral image respectively to obtain a detection map; 2) constructing a background set and an anomaly set based on the detection map; 3) for any pixel to be tested, selecting a specific number of pixels from the background set and the anomaly set to construct a background sub-dictionary and an anomaly sub-dictionary, and merging the two to form a joint dictionary; 4) establishing a collaborative representation model based on the joint dictionary, optimizing and solving it to obtain an estimated coefficient vector; 5) taking the product of the anomaly sub-dictionary and the corresponding coefficient as the detection result. The present invention fully mines the background spectrum and abnormal spectrum information when constructing the dictionary, and effectively avoids the influence of noise in the representation residual vector, thereby improving the performance of hyperspectral anomaly detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral image processing, and further relates to hyperspectral anomaly detection, specifically a hyperspectral anomaly detection method for constructing a joint dictionary guided by spatial-spectral collaboration, which can be used for pest and disease detection, ground object exploration and space exploration. Background Art

[0002] Hyperspectral images are a kind of three-dimensional data that combines images and spectra. The spectral resolution is usually below 10nm and can reach 10 -2 Hyperspectral imagery is on the order of λ, containing continuous spectral information in dozens to hundreds of bands. Depending on whether spectral prior information is included, hyperspectral imagery detection technology can be divided into target detection and anomaly detection. Hyperspectral anomaly detection aims to identify pixels in hyperspectral images whose spectral signals deviate from the background spectrum. Because hyperspectral anomaly detection does not require prior information, it is widely used in fields such as pest and disease detection, land feature exploration, and space exploration.

[0003] To date, collaborative representation-based methods have attracted widespread attention in the field of hyperspectral anomaly detection. Among them, the collaborative representation detector CRD [paper: Collaborative representation for hyperspectral anomaly detection] is a classic method in this field. This method believes that the background pixels to be detected can be linearly represented by the surrounding background pixels (i.e., dictionary), while the abnormal pixels to be detected cannot. Usually, the CRD method uses concentric double windows to select background pixels to construct a background dictionary, and uses the representation residual vector to measure the degree of abnormality of the pixel. However, there are two problems with the above method: 1) In the collaborative representation process, only the background spectral information is usually used, and the abnormal spectral information is ignored, resulting in suboptimal performance; 2) The representation residual vector is usually used to calculate the degree of abnormality of the pixel, but the representation residual vector often contains noise components, which is not conducive to the measurement of abnormal targets. Summary of the Invention

[0004] The present invention aims to address the shortcomings of the above-mentioned prior art and propose a hyperspectral anomaly detection method based on spatial-spectral collaborative guidance of joint dictionary construction, which is used to solve the problem of poor detection performance caused by insufficient mining of background spectrum and abnormal spectrum information and the presence of noise interference when constructing the dictionary. First, the input hyperspectral image is detected using a spatial detector and a spectral detector respectively to obtain a spatial detection map and a spectral detection map; secondly, a background set and an anomaly set are constructed based on the obtained detection maps; for any pixel to be tested, a specific number of pixels are selected from the background set and the anomaly set to construct a background sub-dictionary and an anomaly sub-dictionary, and the two are merged to form a joint dictionary; a collaborative representation model based on the joint dictionary is established and optimized, and the product of the anomaly sub-dictionary and the corresponding representation coefficient is used as the detection result. By combining the background sub-dictionary and the anomaly sub-dictionary, the method of the present invention can fully mine the background spectrum and abnormal spectrum information, and by using the product of the anomaly sub-dictionary and the corresponding representation coefficient as the detection result, it can effectively avoid the influence of the noise component in the representation residual vector on the detection result, thereby improving the hyperspectral anomaly detection result.

[0005] The specific steps of the present invention to achieve the above-mentioned purpose include the following:

[0006] (1) Construct background set and exception set:

[0007] (1.1) Input the hyperspectral image X, and use the spatial detector and spectral detector to detect the hyperspectral image X, and obtain the spatial detection map d spa and spectrum detection diagram d spe ;

[0008] (1.2) Calculate the spatial detection graph d spa and spectrum detection diagram d spe The product of d ae , and use the maximum inter-class variance method to calculate d ae Perform thresholding to obtain a binary image d bin ;

[0009] (1.3) Use the principal component analysis method to reduce the dimension of the hyperspectral image, and use a simple linear iterative clustering method to perform superpixel segmentation on the first three principal components to obtain a superpixel segmentation map Where T represents the total number of superpixel blocks generated, t∈[1,T], n t represents the pixels contained in the t-th superpixel block, Indicates that n is included t The tth superpixel block of pixels; in the binary image d bin In the background set, background pixels are selected and merged to form a background set in Represents the background pixels selected for the vth time, v∈[1,V], V represents the total number of background pixel selections, that is, the total number of pixels contained in the background set, and V<T; select abnormal pixels and merge them to form an abnormal set in represents the qth selected abnormal pixel, q∈[1,Q], Q represents the total number of abnormal pixel selections, that is, the total number of pixels contained in the abnormal set;

[0010] (2) Construct a joint dictionary:

[0011] (2.1) For any pixel x i ∈X,i∈[1,N], where N represents the number of pixels contained in the hyperspectral image, calculate x i With Background Set X B The distance to the first pixel in

[0012]

[0013] Calculate x in turn according to the above distance calculation method i With X B The distance set of all pixels in and will All distances in are arranged in ascending order, and the first hyperparameter is set to k B , according to the first k B The index of the distance from X B Select pixels from the background to construct a background sub-dictionary

[0014] Calculate x in turn according to the above distance calculation method i With exception set X A The distance set of all pixels in and will All distances are arranged in ascending order, and the second hyperparameter is set to k A , according to the first k A The index of the distance from X A Select pixels from the dataset to construct an abnormal sub-dictionary

[0015] (2.2) Merge background sub-dictionary and exception sub-dictionary Forming a joint dictionary

[0016] (3) Constructing a collaborative representation model:

[0017] According to the joint dictionary D i Construct any pixel x i Collaborative representation model:

[0018] xi =D i α i +e i ,

[0019] where x i ∈X,i∈[1,N],α i represents the coefficient vector, e i represents the residual vector;

[0020] (4) The objective function of the collaborative representation model is transformed into α i Find the partial derivative and set it to zero to get the estimated coefficient vector

[0021]

[0022] (5) According to Will Split into estimated background coefficient vector and the estimated anomaly coefficient vector Two parts, using the estimated anomaly coefficient vector Calculate the response value R of the i-th pixel i :

[0023]

[0024] (6) Take i = 1, 2, ..., N in turn to calculate the response values ​​of all pixels in the hyperspectral image X and merge them to obtain the final detection result.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] First, the present invention uses a space-spectrum detector to obtain initial detection results, and uses this to guide the construction of background sets and anomaly sets, thereby establishing a robust joint dictionary. This method can fully mine the background spectrum and anomaly spectrum information, enhance the pixel representation effect, and thus improve the detection performance.

[0027] Second, the method of the present invention can effectively avoid the influence of the noise component in the representation residual vector on the detection result by taking the product of the anomaly sub-dictionary and the corresponding representation coefficient as the detection result, thereby improving the hyperspectral anomaly detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Flowchart for realizing the method of the present invention;

[0029] Figure 2 Schematic diagram of the application of the method of the present invention in a specific example;

[0030] Figure 3 A schematic diagram of the detailed process of constructing the background set and the exception set for the method of the present invention;

[0031] Figure 4 Schematic diagram of different datasets used in the present invention; (a) represents the pseudo-color image and labels of the Salinas dataset; (b) represents the pseudo-color image and labels of the Texas Coast dataset; (c) represents the pseudo-color image and labels of the Gainesville dataset; (d) represents the pseudo-color image and labels of the San Diego dataset; (e) represents the pseudo-color image and labels of the SpecTIR dataset;

[0032] Figure 5 : The figure compares the detection results of the method of the present invention and the mainstream methods on 5 data sets; (Ia) to (If) respectively represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the Salinas data set, and (Ig) represents the label of the Salinas data set; (II-a) to (II-f) respectively represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the Texas Coast data set, and (II-g) represents the label of the Texas Coast data set; (III-a) to (III-f) respectively represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the Gainesville data set, and (III-g) represents the label of the Gainesville data set; (IV-a) to (IV-f) respectively represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the San Diego data set, and (IV-g) represents the label of the San Diego data set. The labels of the Diego dataset; (Va) to (Vf) represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the SpecTIR dataset, respectively; (Vg) represents the label of the SpecTIR dataset;

[0033] Figure 6 Figure 3 is the ROC curve diagram of the method of the present invention and the mainstream method on five data sets; (a), (b), (c), (d), and (e) represent the ROC curves of the method of the present invention and the mainstream method on the Salinas, Texas Coast, Gainesville, San Diego, and SpecTIR data sets, respectively. DETAILED DESCRIPTION

[0034] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0035] Example:

[0036] Reference Figure 1-3 The present invention proposes a hyperspectral anomaly detection method based on spatial-spectral collaborative guidance and joint dictionary construction, which is implemented in the following steps:

[0037] Step 1: Construct background set and exception set:

[0038] (1.1) Input the hyperspectral image X, and use the spatial detector and spectral detector to detect the hyperspectral image X, and obtain the spatial detection map d spa and spectrum detection diagram d spe , as follows:

[0039] Detection using spatial detector: First, the principal component analysis method is used to reduce the dimension of the hyperspectral image X and obtain the first three principal components Then, the difference-fusion method is used to Processing is as follows:

[0040]

[0041] Among them, F r represents the difference-fusion result of the rth principal component, M O (*) and M C (*) represents the morphological opening and closing operations, respectively, and |*| represents the pixel-by-pixel absolute value operation;

[0042] After the difference-fusion process is completed, the obtained F r Perform pixel-by-pixel summation to obtain the initial spatial detection map d init , merge the first three principal components Get pseudo-color image X pca , for d init Perform domain transformation recursive filtering, where X pca As a guide image, d init As the filtering input image, the domain transformation recursive filtering is as follows:

[0043]

[0044] in, Indicates d init The output of the i-th element in , Represents the feedback coefficient, b=|U i -U i-1 |2 represents the distance between adjacent pixels in the transform domain, Indicates d init The i-th element in ;

[0045]

[0046] Among them, U irepresents the i-th pixel in the transform domain, Represents X pca The kth pixel in δ s and δ r are the first smoothing coefficient and the second smoothing coefficient respectively; traverse d init All elements in the , get the spatial detection map d spa ;

[0047] Detection using spectral detector: Input hyperspectral image X, use Calculate the response value of the i-th pixel in the hyperspectral image, where μ and Σ represent the mean vector and covariance matrix of all pixels in the hyperspectral image, respectively. Traverse all pixels in X to obtain the spectral detection map d spe .

[0048] (1.2) Calculate the spatial detection graph d spa and spectrum detection diagram d spe The product of d ae , and use the maximum inter-class variance method to calculate d ae Perform thresholding to obtain a binary image d bin ;

[0049] (1.3) Use the principal component analysis method to reduce the dimension of the hyperspectral image, and use a simple linear iterative clustering method to perform superpixel segmentation on the first three principal components to obtain a superpixel segmentation map Where T represents the total number of superpixel blocks generated, t∈[1,T], n t represents the pixels contained in the t-th superpixel block, Indicates that n is included t The tth superpixel block of pixels; in the binary image d bin In the background set, background pixels are selected and merged to form a background set in Represents the background pixels selected for the vth time, v∈[1,V], V represents the total number of background pixel selections, that is, the total number of pixels contained in the background set, and V<T; select abnormal pixels and merge them to form an abnormal set in represents the qth selected abnormal pixel, q∈[1,Q], Q represents the total number of abnormal pixel selections, that is, the total number of pixels contained in the abnormal set.

[0050] The above binary image d bin In the binary image d, background pixels are selected and merged to form a background set, which is implemented as follows: bin Found in For regions with the same position, determine whether there is a 1 inside the region. If there is no 1, the hyperspectral image pixel corresponding to the center of the superpixel block is selected as the background pixel. If there is a 1, the superpixel block is skipped and no background pixel selection operation is performed. All superpixel blocks are processed in sequence using the above method, and the selected background pixels are merged to form a background set.

[0051] The above binary image d bin In the example, abnormal pixels are selected and merged to form an abnormal set, which is implemented as follows: Find the binary image d bin The index with a value of 1 is selected in the hyperspectral image X according to the index, and the abnormal pixels are merged to form an abnormal set.

[0052] Step 2: Build a joint dictionary:

[0053] (2.1) For any pixel x i ∈X,i∈[1,N], where N represents the number of pixels contained in the hyperspectral image, calculate x i With Background Set X B The distance to the first pixel in

[0054]

[0055] Calculate x in turn according to the above distance calculation method i With X B The distance set of all pixels in and will All distances in are arranged in ascending order, and the first hyperparameter is set to k B , according to the first k B The index of the distance from X B Select pixels from the background to construct a background sub-dictionary

[0056] Calculate x in turn according to the above distance calculation method i With exception set X A The distance set of all pixels in and will All distances are arranged in ascending order, and the second hyperparameter is set to k A , according to the first k A The index of the distance from X A Select pixels from the dataset to construct an abnormal sub-dictionary

[0057] (2.2) Merge background sub-dictionary and exception sub-dictionary Forming a joint dictionary

[0058] Step 3: Build a collaborative representation model:

[0059] According to the joint dictionary D i Construct any pixel x i Collaborative representation model:

[0060] x i =D i α i +e i ,

[0061] where x i ∈X,i∈[1,N],α i represents the coefficient vector, e i represents the residual vector;

[0062] Step 4: The objective function of the collaborative representation model is as follows:

[0063]

[0064] Where β represents the trade-off coefficient, diag(·) indicates the construction of a diagonal matrix, st indicates the constraint condition; [·] T represents the transpose operation, represents the background coefficient vector, represents the outlier coefficient vector.

[0065] The objective function of the collaborative representation model is α i Find the partial derivative and set it to zero to get the estimated coefficient vector

[0066]

[0067] Step 5: According to Will Split into estimated background coefficient vector and the estimated anomaly coefficient vector Two parts, using the estimated anomaly coefficient vector Calculate the response value R of the i-th pixel i :

[0068]

[0069] Step 6: Take i=1, 2, ..., N in turn to calculate the response values ​​of all pixels in the hyperspectral image X and merge them to obtain the final detection result.

[0070] The effects of the present invention will be further described below with reference to experiments.

[0071] 1. Experimental conditions:

[0072] The experiments of the present invention were carried out in a hardware environment with a CPU main frequency of 2.00 GHz, a memory of 16 GB, a Windows 10 operating system, and a software environment of Python 3.6.13.

[0073] 2. Experimental content:

[0074] The datasets used in this experiment include Salinas, Texas Coast, Gainesville, San Diego, and SpecTIR datasets, respectively. Figure 4 For each data set, the detection effect of the method of the present invention is compared with that of five mainstream methods from two perspectives: qualitative (i.e., detection graph and ROC curve) and quantitative (i.e., AUC value).

[0075] The mainstream methods include RX[Paper: Adaptive multiple-band CFAR detection of an optical pattern with unknown spectral distribution], CRD[Paper: Collaborative representation for hyperspectral anomaly detection], CRDBPSW[Paper: Collaborative representation with background purification and saliency weightfor hyperspectral anomaly detection], LSMAD[Paper: A low-rank and sparse matrix decomposition-based mahalanobis distance method for hyperspectral anomaly detection] and KIFD[Paper: Hyperspectral anomaly detection with kernel isolation forest].

[0076] 3. Simulation results and analysis:

[0077] Figure 5The following is a comparison chart of the detection results of the proposed method and mainstream methods on five datasets; (I)-(V) correspond to the Salinas, Texas Coast, Gainesville, San Diego, and SpecTIR datasets, respectively; (a)-(e) correspond to the five existing methods, RX, CRD, CRDBPSW, LSMAD, and KIFD, respectively; (f) corresponds to the proposed method; and (g) corresponds to the label; see specifically: Figure 5 In the figure, (Ia) to (If) represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the Salinas dataset, respectively, and (Ig) represents the label of the Salinas dataset; (II-a) to (II-f) represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the Texas Coast dataset, respectively, and (II-g) represents the label of the Texas Coast dataset; (III-a) to (III-f) represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the Gainesville dataset, respectively, and (III-g) represents the label of the Gainesville dataset; (IV-a) to (IV-f) represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the San Diego dataset, respectively, and (IV-g) represents the label of the San Diego dataset. The labels of the Diego dataset; (Va) to (Vf) respectively represent the detection results of RX, CRD, CRDBPSW, LSMAD, KIFD and the method of the present invention on the SpecTIR dataset, and (Vg) represents the label of the SpecTIR dataset.

[0078] By comparing the existing method, the method of the present invention and the label, it can be seen that the method of the present invention has the best effect and is significantly better than the mainstream methods in the existing five prior arts.

[0079] Figure 6 Figure 3 is the ROC curve of the method of the present invention and the mainstream method on five data sets; (a) represents the Salinas data set, (b) represents the Texas Coast data set, (c) represents the Gainesville data set, (d) represents the San Diego data set, and (e) represents the SpecTIR data set.

[0080] The larger the area under the ROC curve, the better the detection performance. Figure 6 It can be seen that the detection effect of the method of the present invention is the best.

[0081] The following table (Table 1) lists the AUC values ​​of the method of the present invention and five mainstream methods on five data sets. The AUC value is between 0 and 1, and the closer the value is to 1, the better the detection performance.

[0082] Table 1 Comparison of AUC values ​​of the proposed method and five mainstream methods on five data sets

[0083] Dataset RX CRD CRDBPSW LSMAD KIFD The present invention Salinas 0.8073 0.9635 0.9932 0.9375 0.9935 0.9983 Texas Coast 0.9907 0.9796 0.9956 0.9928 0.9357 0.9988 Gainesville 0.9597 0.9697 0.9522 0.9645 0.9917 0.9933 San Diego 0.9403 0.9412 0.9862 0.9701 0.9919 0.9941 SpecTIR 0.9914 0.9920 0.9991 0.9984 0.9969 0.9997

[0084] As can be seen from the above table, on five different data sets (Salinas, Texas Coast, Gainesville, San Diego and SpecTIR data sets), the method of the present invention has the largest AUC value, further indicating that the method of the present invention has the best performance compared with existing methods.

[0085] The above experimental results prove the correctness and effectiveness of the method proposed in the present invention.

[0086] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A hyperspectral anomaly detection method based on spatial-spectral collaborative guidance and joint dictionary construction, characterized in that: The specific steps are as follows: (1) Construct background set and exception set: (1.1) Input the hyperspectral image X, and use the spatial detector and spectral detector to detect the hyperspectral image X, and obtain the spatial detection map d spa and spectrum detection diagram d spe ; (1.2) Calculate the spatial detection graph d spa and spectrum detection diagram d spe The product of d ae , and use the maximum inter-class variance method to calculate d ae Perform thresholding to obtain a binary image d bin ; (1.3) Use the principal component analysis method to reduce the dimension of the hyperspectral image, and use a simple linear iterative clustering method to perform superpixel segmentation on the first three principal components to obtain a superpixel segmentation map Where T represents the total number of superpixel blocks generated, t∈[1,T], n t represents the pixels contained in the t-th superpixel block, Indicates that n is included t The tth superpixel block of pixels; in the binary image d bin In the background set, background pixels are selected and merged to form a background set in Represents the background pixels selected for the vth time, v∈[1,V], V represents the total number of background pixel selections, that is, the total number of pixels contained in the background set, and V<T; select abnormal pixels and merge them to form an abnormal set in represents the qth selected abnormal pixel, q∈[1,Q], Q represents the total number of abnormal pixel selections, that is, the total number of pixels contained in the abnormal set; (2) Construct a joint dictionary: (2.1) For any pixel x i ∈X,i∈[1,N], where N represents the number of pixels contained in the hyperspectral image, calculate x i With Background Set X B The distance to the first pixel in Calculate x in turn according to the above distance calculation method i With X B The distance set of all pixels in and will All distances in are arranged in ascending order, and the first hyperparameter is set to k B , according to the first k B The index of the distance from X B Select pixels from the background to construct a background sub-dictionary Calculate x in turn according to the above distance calculation method i With exception set X A The distance set of all pixels in And d i A All distances are arranged in ascending order, and the second hyperparameter is set to k A , according to the first k A The index of the distance from X A Select pixels from the dataset to construct an abnormal sub-dictionary (2.2) Merge background sub-dictionary and exception sub-dictionary Forming a joint dictionary (3) Constructing a collaborative representation model: According to the joint dictionary D i Construct any pixel x i Collaborative representation model: x i =D i α i +e i , where x i ∈X,i∈[1,N],α i represents the coefficient vector, e i represents the residual vector; (4) The objective function of the collaborative representation model is transformed into α i Find the partial derivative and set it to zero to get the estimated coefficient vector (5) According to Will Split into estimated background coefficient vector and the estimated anomaly coefficient vector Two parts, using the estimated anomaly coefficient vector Calculate the response value R of the i-th pixel i : (6) Take i = 1, 2, ..., N in turn to calculate the response values ​​of all pixels in the hyperspectral image X and merge them to obtain the final detection result.

2. The method according to claim 1, wherein: In step (1), the spatial detector and the spectral detector are used to detect the hyperspectral image X, respectively, as follows: Detection using spatial detector: First, the principal component analysis method is used to reduce the dimension of the hyperspectral image X and obtain the first three principal components Then, the difference-fusion method is used to Processing is as follows: Among them, F r represents the difference-fusion result of the rth principal component, M O (*) and M C (*) represents the morphological opening and closing operations, respectively, and * represents the pixel-by-pixel absolute value operation; After the difference-fusion process is completed, the obtained F r Perform pixel-by-pixel summation to obtain the initial spatial detection map d init , merge the first three principal components Get pseudo-color image X pca , for d init Perform domain transformation recursive filtering, where X pca As a guide image, d init As the filtering input image, the domain transformation recursive filtering is as follows: in, Indicates d init The output of the i-th element in , Indicates the feedback coefficient, b=U i -U i-12 represents the distance between adjacent pixels in the transform domain, Indicates d init The i-th element in ; Among them, U i represents the i-th pixel in the transform domain, Represents X pca The kth pixel in δ s and δ r are the first smoothing coefficient and the second smoothing coefficient respectively; traverse d init All elements in the , get the spatial detection map d spa ; Detection using spectral detector: Input hyperspectral image X, use i∈[1,N] calculates the response value of the i-th pixel in the hyperspectral image, where μ and Σ represent the mean vector and covariance matrix of all pixels in the hyperspectral image, respectively. Traverse all pixels in X to obtain the spectral detection map d spe .

3. The method according to claim 1, wherein: In step (1.3), the binary image d bin In the binary image d, background pixels are selected and merged to form a background set, which is implemented as follows: bin Found in For regions with the same position, determine whether there is a 1 inside the region. If there is no 1, the hyperspectral image pixel corresponding to the center of the superpixel block is selected as the background pixel. If there is a 1, the superpixel block is skipped and no background pixel selection operation is performed. All superpixel blocks are processed in sequence using the above method, and the selected background pixels are merged to form a background set.

4. The method according to claim 1, wherein: In step (1.3), the binary image d bin In the example, abnormal pixels are selected and merged to form an abnormal set, which is implemented as follows: Find the binary image d bin The index with a value of 1 is selected in the hyperspectral image X according to the index, and the abnormal pixels are merged to form an abnormal set.

5. The method according to claim 1, wherein: The objective function of the collaborative representation model in step (4) is as follows: Where β represents the trade-off coefficient, diag(·) indicates the construction of a diagonal matrix, st indicates the constraint condition; [·] T represents the transpose operation, represents the background coefficient vector, represents the outlier coefficient vector.

Images