A fast and robust method and system for abnormal target detection in hyperspectral images

Abstract

The present invention belongs to the technical field of remote sensing and target detection, and discloses a method and system for detecting abnormal targets in hyperspectral images with fast and strong robustness. The method comprises: selecting characteristic bands of hyperspectral images using an axis-parallel subspace method; dividing q characteristic bands into N groups, and calculating Mahalanobis distance for each group of data; solving the abnormal score map based on frequency histogram corresponding to the Mahalanobis distance map of N groups of data; fusing the abnormal score of the Mahalanobis distance map and frequency histogram of each grouped data; calculating the binary accumulation result of N groups of data, and realizing abnormal target detection through energy threshold. The present invention realizes a fast, accurate and robust abnormal detection method, which can be applied to the rapid processing of abnormal target detection in hyperspectral images, making up for the shortcomings of high computational complexity and slow detection speed of existing high-precision abnormal target detection methods, and providing a basis for the wide application and identification of real-time processing of hyperspectral images.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing and target detection, and in particular relates to a fast and robust hyperspectral image abnormal target detection method and system. Background Art

[0002] At present, hyperspectral images are a kind of three-dimensional data block, which has extremely rich spectral information in addition to spatial information. Therefore, in the past few decades, hyperspectral images have been widely used and developed in the fields of target detection, recognition, and remote sensing. In the field of target detection, the detection algorithm is divided into two categories: supervised and unsupervised detection according to whether the target prior information is available. In practical applications, the background and target prior information of hyperspectral images are difficult to obtain, so anomaly detection in unsupervised detection is widely used. Anomaly detection usually uses the spatial anomaly and spectral anomaly characteristics of the target to detect abnormal targets, that is, the spectral anomaly characteristics are manifested as the spectral information of the abnormal target and the background of the hyperspectral image is extremely different; the spatial anomaly characteristics are manifested as the abnormal target is sparsely distributed in the entire hyperspectral image, and it occupies a small number of pixels and has a low probability of occurrence.

[0003] The existing methods for detecting abnormal targets in hyperspectral images are mainly divided into three categories. On the one hand, they use spectral anomaly information. Assuming that the background information obeys Gaussian distribution, the abnormal targets are determined by calculating the Mahalanobis distance. Since the background information of the hyperspectral image does not necessarily strictly obey Gaussian distribution, the detection effect of this type of method is general, but the algorithm complexity is small and the detection speed is fast. The second type is based on the spectral anomaly characteristics of representation. The core idea is that the background in the hyperspectral image can be represented by its surrounding pixels, while the abnormal targets cannot be represented by the surrounding pixels. The detection accuracy of this type of algorithm is greatly improved, but the computational complexity is also increased accordingly, and the real-time performance of hyperspectral image anomaly detection is poor. The third type of method is to use spatial anomaly characteristics, assuming that the abnormal targets are sparsely distributed, and use isolation forests to find abnormal targets. However, this type of algorithm has good detection effect when the background is simple and clean, and the detection accuracy decreases when the background is complex, and the algorithm robustness is poor.

[0004] Existing methods for detecting abnormal targets in hyperspectral images are already very mature. However, due to the lack of prior information about targets and backgrounds and the large amount of hyperspectral image data, anomaly detection algorithms with good detection effects and strong robustness usually require a lot of calculations, and their detection time is long and it is difficult to ensure real-time performance. Anomaly detection algorithms with good real-time performance have poor detection accuracy and weak robustness. The detection effect is good only when the background of the hyperspectral image is simple and clean, but poor when the background is complex.

[0005] Therefore, in the field of anomaly detection with short detection time and good detection effect, the existing hyperspectral image anomaly detection algorithm cannot meet the demand. The present invention has not found any corresponding literature and reports on fast and robust hyperspectral image anomaly detection methods in the relevant scope of search and new search. Therefore, there is an objective need for a fast, strong and robust anomaly detection algorithm with high detection accuracy that is suitable for practical applications.

[0006] Through the above analysis, the problems and defects of the prior art are as follows:

[0007] (1) The existing methods for detecting abnormal targets in hyperspectral images have high computational complexity, long detection time, and poor real-time performance. Moreover, when the background is complex, the detection accuracy decreases and the algorithm has poor robustness.

[0008] (2) In the field of anomaly detection with short detection time and good detection effect, the existing hyperspectral image anomaly detection algorithms cannot meet the needs.

[0009] (3) No corresponding literature or reports on fast and robust hyperspectral image anomaly detection methods were found through searching and checking within the relevant scope.

[0010] The difficulty of solving the above problems and defects is:

[0011] Anomaly detection is applied in the field of target detection when there is a lack of hyperspectral image background and target information. Due to the lack of prior knowledge, the robustness of existing anomaly detection algorithms is poor. The detection accuracy is different for different hyperspectral images, and the higher the detection accuracy, the greater the computational complexity of the algorithm. It is challenging to design anomaly detection algorithms that improve accuracy and robustness without increasing algorithm complexity.

[0012] The significance of solving the above problems and defects is:

[0013] Designing a hyperspectral image anomaly detection algorithm with high detection accuracy and strong robustness without increasing the complexity of the algorithm can serve the development of the field of real-time processing of abnormal targets in hyperspectral images, improve the detection performance of abnormal target detection in practical applications, and respond quickly to moving abnormal targets, thus promoting the application of hyperspectral image target detection in industry. Summary of the invention

[0014] In view of the problems existing in the prior art, the present invention provides a fast and robust method and system for abnormal target detection in hyperspectral images.

[0015] The present invention is implemented in this way: a fast and robust hyperspectral image abnormal target detection method, the fast and robust hyperspectral image abnormal target detection method comprising the following steps:

[0016] Step 1: Use the axis-parallel subspace method to select the characteristic bands of the hyperspectral image;

[0017] The axis-parallel subspace method has low computational complexity and is suitable for the field of anomaly detection of high-dimensional data. Using it to select the characteristic bands of highlight images can effectively retain the abnormal information in the hyperspectral image and eliminate the spectral data redundancy between bands, thereby reducing the input data dimension of the anomaly detection algorithm and reducing the computational complexity of the anomaly detection algorithm.

[0018] Step 2: divide the selected q feature bands into N groups, and calculate the Mahalanobis distance for each group of data;

[0019] If the selected feature bands are grouped, the Mahalanobis distance calculation of each group can be processed in parallel on the hardware, further saving the detection time.

[0020] Step 3, solving the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data;

[0021] The Mahalanobis distance map is a calculation result based on the spectral anomaly characteristics, that is, the spectral characteristics of the abnormal target in the hyperspectral image and the spectral characteristics of the background are greatly different as the detection basis information. The anomaly score based on the frequency histogram is based on the spatial anomaly characteristics, that is, the abnormal target in the hyperspectral image is sparse in space relative to the entire hyperspectral image.

[0022] Step 4: Fusion the abnormality scores of the Mahalanobis distance graph and the frequency histogram of each grouped data;

[0023] This process fuses the Mahalanobis distance map obtained based on the spectral anomaly characteristics and the frequency histogram obtained based on the spatial anomaly characteristics, making full use of the spatial-spectral information of the hyperspectral image and improving the detection accuracy of abnormal targets.

[0024] Step 5: Calculate the binary accumulation results of N groups of data and realize abnormal target detection through energy threshold.

[0025] Binary accumulation is performed on the detection results of N groups of images. It is assumed that the false detection noise appears randomly in the detection results of each group, while the position of the real abnormal target is fixed. Through binary accumulation, the energy of the position of the abnormal target is much greater than that of the noise. This improves the robustness of the anomaly detection algorithm.

[0026] Furthermore, in step 1, the method of using the axis-parallel subspace to select the characteristic bands of the hyperspectral image includes:

[0027] Hyperspectral image is represented as X∈R m×n×l; where m and n represent the length and width of each band image, and l represents the number of bands in the hyperspectral image; the value of the abnormal target is greatly different from the background value and is sparsely distributed, so the more concentrated the background pixel value is, the easier it is to detect the abnormal target away from the background. Based on this cognition, the data variance var of each characteristic band image is calculated j , sorted by variance, and starting from the band with the smallest variance, q feature bands are selected as the subspace for abnormal target detection.

[0028] Among them, the data variance var of each characteristic band image j Calculations include:

[0029] (1) Calculate the mean u of each feature band image j : j represents the jth characteristic band, x ab represents the i-th pixel of the j-th feature band, then the mean value u of each feature band image is determined j for:

[0030]

[0031] (2) Calculate the data variance var of each characteristic band image j :Where dist represents the band data and the mean u of the band j The distance between them, Card(X) is the matrix basis of the hyperspectral image, and the data variance of each characteristic band image is expressed as:

[0032]

[0033] Furthermore, in step 2, the selected q characteristic bands are evenly divided into N groups, and the Mahalanobis distance is calculated for each group of data, including:

[0034] Input the selected q characteristic band data X sub ∈R m×n×q , the data is evenly divided into N groups, the i-th group is represented by: X sub_i ∈R m×n×q / N ; Calculate the Mahalanobis distance D of the i-th group of subspace feature data Ma_i .

[0035] The Mahalanobis distance D corresponding to the N sets of characteristic band data is calculated. Ma_i ,include:

[0036] (1) Calculate the covariance matrix Γ of the i-th group of characteristic band data Xsub_i :Input the characteristic band data X of the i-th group sub_i ∈R m×n×q / N , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i _2D∈Rq / N×mn , the covariance matrix Γ of the i-th group of characteristic band data Xsub_i ∈R q / N×q / N It is expressed as:

[0037] Γ Xsub_i =X sub_i _2D×X sub_i _2D T ;

[0038] (2) Calculate the mean matrix u of the i-th group of characteristic band data X_i :Input the characteristic band data X of the i-th group sub_i ∈R m ×n×q / N , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i _2D∈R q / N×mn , the mean matrix u of the i-th group of characteristic band data X_i ∈R q / N×1 It is expressed as:

[0039]

[0040] (3) Calculate the Mahalanobis distance D corresponding to N sets of characteristic band data Ma_i : Input the covariance matrix Γ of the i-th group of feature band data Xsub_i ∈R q / N×q / N , input the mean matrix u of the i-th group of feature band data X ∈R q / N×1 , the Mahalanobis distance D corresponding to the characteristic band data of the i-th group Ma_i_1 ∈R mn×1 :

[0041] Where a=1, 2, ... mn;

[0042] D Ma_i_1 The Mahalanobis distance graph D corresponding to the final i-th group of feature data is obtained by converting the mn×1-dimensional matrix into an m×n two-dimensional matrix. Ma_i ∈R m×n .

[0043] Further, in step 3, the abnormal score graph S based on the frequency histogram corresponding to the Mahalanobis distance graph of the N groups of data is solved hist_i ,include:

[0044] Input the Mahalanobis distance graph D of group i Ma_i , the number of groups k in the frequency histogram; convert the data in the Mahalanobis distance map into a one-dimensional array and sort D Mai_1d, fix the amount of data contained in each group to N_fix, calculate the starting value and cutoff value of each group of data in the frequency histogram, and obtain the width of each group. Then the height of each group in the frequency histogram histogram (k) can be obtained by N_fix and the width of each group. Finally, the abnormal score of each group based on the frequency histogram is S hist_i ,include:

[0045] Input the Mahalanobis distance graph D of N sets of data Ma_1 , D Ma_2 , …D Ma_N , input the hist(k) obtained by the Mahalanobis distance map of the i-th group, then the abnormal score result S based on the frequency histogram corresponding to the pixel points of the k-th group hist_i It is expressed as:

[0046]

[0047] S hist_i_1 The final abnormal score result S based on the frequency histogram corresponding to the i-th group of feature data is obtained by converting the mn×1-dimensional matrix into an m×n two-dimensional matrix. hist_i ∈R m×n .

[0048] Further, in step 4, the abnormality score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused, including inputting the Mahalanobis distance map D Ma_i and the anomaly score map S based on the frequency histogram hist_i , the fusion result is S total_i ;

[0049] The abnormal score fusion result S of the Mahalanobis distance graph and the frequency histogram of each grouped data is total_i Solving, including:

[0050] Input the i-th group of anomaly score results S based on frequency histogram hist_i ∈R m×n , input the Mahalanobis distance graph D corresponding to the i-th group of feature data Ma_i ∈R m×n , then the anomaly score fusion result S of the i-th group total_i It is expressed as:

[0051]

[0052] Among them, λ S and λ D is the fusion coefficient.

[0053] Furthermore, in step 5, the binary accumulation results of the N groups of data are calculated, and abnormal target detection is achieved through energy thresholds, including:

[0054] (1) Calculate the binary accumulation of N sets of data B(a, b)∈R m×n Results: Fusion anomaly score map S of input N groups of data total_i , set the threshold Val threshold , greater than Val threshold The pixel position is considered to be the location of the abnormal target, which is less than Val threshold The pixel position of each group is considered as the background, and the abnormal target pixel position of each group is assigned 1, and the other background pixels are assigned 0, and the abnormal detection results of N groups of data are accumulated;

[0055] (2) Determine the anomaly detection target: Input the binary accumulation result of N groups of data, design the energy threshold, and consider the pixels with energy greater than the threshold as anomaly targets, and the pixels with energy less than the threshold as background.

[0056] Wherein, the binary accumulation B(a, b)∈R of the N sets of data is calculated m×n Results, including:

[0057] 1) Input the anomaly score fusion result S of N groups of data total_1 , S total_2 ,…S total_N , input threshold value Val threshold , the binary calculation result diagram of the i-th group is M total_i (a, b); (a, b) represents the coordinate position of the pixel, i represents the i-th group of characteristic band data; M total_i The calculation process is as follows:

[0058]

[0059] 2) Input N sets of data and the binary calculation result graph M total_1 , M total_2 , …M total_N , the final binary accumulation of N sets of data B(a, b)∈R m×n The results are expressed as:

[0060]

[0061] Another object of the present invention is to provide a fast and robust hyperspectral image abnormal target detection system using the fast and robust hyperspectral image abnormal target detection method, the fast and robust hyperspectral image abnormal target detection system comprising:

[0062] The feature band selection module is used to select the feature bands of the hyperspectral image using the axis-parallel subspace method;

[0063] A Mahalanobis distance calculation module is used to divide the selected q characteristic bands into N groups and calculate the Mahalanobis distance for each group of data;

[0064] An abnormal score map acquisition module is used to solve the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of N groups of data;

[0065] The anomaly score fusion module is used to fuse the anomaly scores of the Mahalanobis distance map and frequency histogram of each grouped data;

[0066] The abnormal target detection module is used to calculate the binary accumulation results of N groups of data and realize abnormal target detection through energy threshold.

[0067] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0068] The axis-parallel subspace method is used to select the characteristic bands of the hyperspectral image; the selected q characteristic bands are divided into N groups, and the Mahalanobis distance is calculated for each group of data; the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data is solved; the abnormal score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused; the binary accumulation results of N groups of data are calculated, and abnormal target detection is achieved through energy threshold.

[0069] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:

[0070] The axis-parallel subspace method is used to select the characteristic bands of the hyperspectral image; the selected q characteristic bands are divided into N groups, and the Mahalanobis distance is calculated for each group of data; the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data is solved; the abnormal score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused; the binary accumulation results of N groups of data are calculated, and abnormal target detection is achieved through energy threshold.

[0071] Another object of the present invention is to provide an information data processing terminal, which is used to implement the fast and robust hyperspectral image abnormal target detection system.

[0072] In combination with all the above-mentioned technical solutions, the advantages and positive effects possessed by the present invention are as follows: The fast and highly robust method for abnormal target detection in hyperspectral images provided by the present invention mainly relates to abnormal target detection in hyperspectral images, and is used for the rapid processing of abnormal target detection in remote sensing. The present invention solves the defects that the detection accuracy and detection complexity cannot be taken into account at the same time due to the large amount of hyperspectral image data and the lack of target prior information, and the detection method has poor robustness, and realizes a fast, accurate and robust abnormality detection method. At the same time, the present invention can be applied to the rapid processing of abnormal target detection in hyperspectral images, which makes up for the shortcomings of the existing high-precision abnormal target detection methods with high computational complexity and slow detection speed, and lays a foundation for the identification of the widespread application of real-time processing of hyperspectral images. The present invention also includes the following beneficial effects:

[0073] (1) Fully utilizing the spectral anomaly characteristics and spatial anomaly characteristics: When designing the anomaly scoring strategy, the present invention integrates the anomaly scoring strategy based on the spectral anomaly characteristics and the spatial anomaly characteristics, that is, utilizing the difference between the spectrum of the abnormal target and the background image and combining the characteristics that the abnormal target is spatially dispersed and sparse. The accuracy of abnormal target detection is improved. Compared with the existing anomaly detection algorithm, the algorithm complexity does not increase while achieving the improvement of detection accuracy by combining the spatial spectrum.

[0074] (2) The model uses subspace grouping and binary accumulation methods to further improve algorithm efficiency and enhance robustness: In view of the data redundancy of hyperspectral images, the present invention uses an axis-parallel subspace algorithm to select subspectral bands to reduce data redundancy. By grouping the selected subspaces, the effect of parallel processing of each group of data is achieved, which greatly improves the detection efficiency of the anomaly detection method. It is believed that in the detection results of each subspace, the position of the real abnormal target is fixed and the position of the false detection appears randomly. Binary accumulation is used to further improve the robustness of the algorithm. A fast and robust hyperspectral image anomaly detection algorithm is realized. It lays the foundation for the application of hyperspectral images in the field of rapid target detection and recognition.

[0075] The present invention provides a fast and robust method for detecting abnormal targets in hyperspectral images. By combining the spectral anomaly information and spatial anomaly information of the abnormal targets, the accuracy of anomaly detection is improved. The detection speed and the robustness of the enhanced detection algorithm are improved based on subspace grouping and binary accumulation, respectively. The present invention provides new prospects for the industrial application of hyperspectral images due to computing power limitations.

[0076] The invention discloses a method for detecting abnormal targets in a hyperspectral image with high robustness, which mainly solves the defects that the detection accuracy and detection complexity cannot be taken into account at the same time and the detection method has poor robustness due to the large amount of hyperspectral image data and the lack of target prior information, and realizes a method for detecting abnormal targets with high robustness. The implementation process is: using the axis-parallel subspace method to select the characteristic bands of the hyperspectral image; the selected q characteristic bands are divided into N groups, and the Mahalanobis distance is calculated for each group of data; the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data is solved; the abnormal score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused; the binary accumulation results of the N groups of data are calculated, and abnormal target detection is realized through the energy threshold. The invention can be applied to the rapid processing of abnormal target detection in hyperspectral images, which makes up for the shortcomings of the existing high-precision abnormal target detection methods with high computational complexity and slow detection speed, and lays a foundation for the identification of the wide application of real-time processing of hyperspectral images. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0078] Figure 1 It is a schematic diagram of a fast and robust method for abnormal target detection in hyperspectral images provided by an embodiment of the present invention.

[0079] FIG. 2 is a schematic diagram of an axis-parallel subspace method provided by an embodiment of the present invention.

[0080] Figure 3 It is a principle diagram of an abnormality scoring strategy based on a frequency histogram provided in an embodiment of the present invention.

[0081] Figure 4 It is an anomaly scoring strategy diagram combining spectral anomaly characteristics and spatial anomaly characteristics provided by an embodiment of the present invention.

[0082] Figure 5 It is a schematic diagram of the location of a real abnormal target provided by an embodiment of the present invention.

[0083] Figure 6 It is a Mahalanobis distance anomaly score graph of 20 groups of subspaces provided by an embodiment of the present invention.

[0084] Figure 7 It is an abnormal score graph based on a frequency histogram of 20 groups of subspaces provided by an embodiment of the present invention.

[0085] Figure 8This is an abnormal result diagram after fusion of 20 groups of subspaces provided by an embodiment of the present invention.

[0086] Fig. 9 It is a binary result diagram provided by an embodiment of the present invention when the threshold value Valthreshold=0.5.

[0087] Fig.10 It is a schematic diagram of the final detection result after binary accumulation provided by an embodiment of the present invention.

[0088] Fig.11 It is a flow chart of a fast and robust method for abnormal target detection in hyperspectral images provided by an embodiment of the present invention.

[0089] Fig.12 It is a structural block diagram of a fast and robust hyperspectral image abnormal target detection system provided by an embodiment of the present invention;

[0090] In the figure: 1. Feature band selection module; 2. Mahalanobis distance calculation module; 3. Anomaly score map acquisition module; 4. Anomaly score fusion module; 5. Abnormal target detection module. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0092] In view of the problems existing in the prior art, the present invention provides a fast and robust method and system for abnormal target detection in hyperspectral images. The present invention is described in detail below in conjunction with the accompanying drawings.

[0093] like Fig.11 As shown, the fast and robust hyperspectral image abnormal target detection method provided by the embodiment of the present invention includes the following steps:

[0094] S101, using the axis-parallel subspace method to select the characteristic bands of the hyperspectral image;

[0095] S102, dividing the selected q characteristic bands into N groups, and calculating the Mahalanobis distance for each group of data;

[0096] S103, solving the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data;

[0097] S104, fusing the abnormality scores of the Mahalanobis distance graph and the frequency histogram of each grouped data;

[0098] S105, calculating the binary accumulation results of N groups of data, and realizing abnormal target detection through energy threshold.

[0099] like Fig.12 As shown, the fast and robust hyperspectral image abnormal target detection system provided by the embodiment of the present invention includes:

[0100] The feature band selection module 1 is used to select the feature band of the hyperspectral image using the axis-parallel subspace method;

[0101] A Mahalanobis distance calculation module 2 is used to divide the selected q characteristic bands into N groups and calculate the Mahalanobis distance for each group of data;

[0102] Anomaly score map acquisition module 3, used to solve the anomaly score map based on the frequency histogram corresponding to the Mahalanobis distance map of N groups of data;

[0103] Anomaly score fusion module 4, used to fuse the anomaly scores of the Mahalanobis distance graph and the frequency histogram of each grouped data;

[0104] The abnormal target detection module 5 is used to calculate the binary accumulation results of N groups of data and realize abnormal target detection through energy threshold.

[0105] The technical solution of the present invention is further described below in conjunction with embodiments.

[0106] Example 1

[0107] At present, various abnormal target detection methods related to the situation that it is difficult to obtain prior information on both the target and the background in hyperspectral images already exist. However, due to the large amount of hyperspectral image data, the algorithm complexity of the high-precision and strong robustness anomaly detection method is very high, and it takes a long time to process, which makes it difficult to meet the application scenarios with high real-time requirements. The existing fast anomaly detection algorithms have low detection accuracy or are only suitable for anomaly detection of hyperspectral images with simple backgrounds. Changes in the background have a great impact on the detection results, and the performance of the detection method is unstable. In view of the shortcomings of the existing algorithms, the present invention makes full use of the spectral anomaly and spatial anomaly characteristics of hyperspectral images, and proposes a fast and strong robust hyperspectral image abnormal target detection method based on subspace grouping and binary accumulation.

[0108] The present invention is a fast and robust method for detecting abnormal targets in hyperspectral images. Figure 1 ,The fast and robust hyperspectral image anomaly detection method includes the following steps:

[0109] S1: Using the axis-parallel subspace method to select the characteristic bands of the hyperspectral image: The hyperspectral image can be represented as X∈R m×n×l, where m and n represent the length and width of each band image, and l represents the number of bands in the hyperspectral image. It is generally believed that the value of the abnormal target is greatly different from the background value and is sparsely distributed. Therefore, the more concentrated the background pixel value is, the easier it is to detect the abnormal target away from the background, as shown in Figure 2. Based on this assumption, the data variance var of each feature band image is calculated. j , sorted by variance, and starting from the band with the smallest variance, q feature bands are selected as the subspace for abnormal target detection.

[0110] The axis-parallel subspace method is a method for selecting characteristic bands of hyperspectral images suitable for anomaly detection. The axis-parallel subspace method can greatly reduce the band redundancy of hyperspectral images and find relevant bands suitable for anomaly detection. The band selection method has low computational complexity and occupies almost no computing resources, which reduces the amount of data for subsequent abnormal target detection and improves the anomaly detection performance.

[0111] S2: Divide the selected q feature bands into N groups and calculate the Mahalanobis distance for each group of data: Input the selected q feature band data X sub ∈R m×n×q , and evenly divide it into N groups: X sub_i ∈R m×n×q / N . Calculate the Mahalanobis distance D of the i-th group of subspace feature data Ma_i .

[0112] The specific expression of the Mahalanobis distance of the i-th group is as follows:

[0113]

[0114] where Γ Xsub_i is the covariance matrix of the i-th group of characteristic band data, u X_i is the mean matrix of the i-th group of characteristic band data, x a is the pixel data to be detected.

[0115] The present invention realizes parallel calculation of multiple groups of data by grouping subspaces, thereby reducing computational complexity and improving detection efficiency. The present invention utilizes the spectral anomaly characteristics of abnormal targets, that is, background targets all obey Gaussian distribution, and there is a great difference in the distribution of abnormal targets and background targets. The larger the calculation result of the Mahalanobis distance, the more likely the pixel point is an abnormal target. However, since the spatial anomaly information of the abnormal target is not used, the false detection rate is usually very high when the abnormal target is determined only based on the Mahalanobis distance.

[0116] S3: Solve the abnormal score map S based on the frequency histogram corresponding to the Mahalanobis distance map of N groups of data hist_i :Input the Mahalanobis distance graph D of the i-th group Ma_i, the number of groups k of the frequency histogram. Convert the data in the Mahalanobis distance map into a one-dimensional array and sort it D Mai _1d, fix the amount of data contained in each group N_fix, such as Figure 3 As shown, the starting value and cutoff value of each group of data in the frequency histogram are calculated to obtain the width of each group. Then the height of each group in the frequency histogram hist(k) can be obtained by N_fix and the width of each group. Finally, the abnormal score of each group based on the frequency histogram is S hist_i .

[0117] The anomaly score results for each group based on the frequency histogram are:

[0118]

[0119] The present invention assumes that the pixels with similar anomaly scores calculated by the Mahalanobis distance are of the same material. Therefore, the frequency histogram distribution is obtained based on the Mahalanobis distance anomaly score graph. When the number of pixels in each group of the fixed frequency histogram is the same, the higher the frequency distribution height of each group of data, the denser the distribution of the group of data, the less likely it is an abnormal target. Conversely, the sparser the distribution, the greater the possibility of an abnormal target. Therefore, the anomaly scoring strategy based on the frequency histogram makes full use of the spatial anomaly characteristics of the sparser spatial distribution of abnormal targets in hyperspectral images.

[0120] S4: Fusion of the abnormality scores of the Mahalanobis distance map and frequency histogram of each grouped data: Figure 4 As shown, input the Mahalanobis distance graph D Ma_i and the anomaly score map S based on the frequency histogram hist_i , the fusion result is S total_i The solution method for the fused anomaly score map is as follows:

[0121]

[0122] where λ S and λ D is the fusion coefficient.

[0123] The present invention obtains a total anomaly score map by fusing the Mahalanobis distance anomaly score map and the anomaly score map based on the frequency histogram. The total anomaly score map fully considers the spectral anomaly characteristics and spatial anomaly characteristics of abnormal targets in the hyperspectral image, further improving the accuracy of abnormal target detection.

[0124] S5: Calculate the binary accumulation of N sets of data B(a, b)∈R m×n Results: Fusion anomaly score map S of input N groups of data total_i , set the threshold Val threshold , greater than Val thresholdThe pixel position is considered to be the location of the abnormal target, which is less than Val threshold The pixel positions of the target abnormalities are considered as background, and the abnormal target pixel positions of each group are assigned 1, and the other background pixels are assigned 0. The abnormality detection results of N groups of data are accumulated.

[0125] The present invention considers that the position of the real abnormal target in each set of data in the hyperspectral image is fixed, while the position of the virtual abnormal target appears randomly in each set of total abnormal detection result graphs. Through the binary accumulation of the detection results of each set of data, the energy of the position of the real abnormal target will become larger and larger, while the random accumulation of the energy of the virtual detection position is very different from the energy of the real abnormal target. The binary accumulation of the present invention can effectively improve the stability of the detection performance and enhance the robustness of the abnormal detection algorithm.

[0126] S6: Determine the abnormal detection target: Input the binary accumulation result of N groups of data, design the energy threshold, and the pixel position with a value greater than the energy threshold is considered to be an abnormal target, and the pixel position with a value less than the energy threshold is considered to be the background.

[0127] The present invention realizes a fast and robust method for abnormal target detection in hyperspectral images. Compared with the existing abnormal target detection methods in hyperspectral images, the present invention makes full use of the spectral anomaly characteristics and spatial anomaly characteristics of abnormal targets, and uses subspace grouping and binary accumulation to enable parallel processing of each group of subspace anomaly detection, thereby improving the detection accuracy of anomaly detection while reducing the detection time, laying a foundation for the rapid response of hyperspectral images in the field of target detection and recognition.

[0128] Example 2

[0129] The method for detecting abnormal targets in a fast and robust hyperspectral image of the present invention is the same as that in Example 1, except that the step S1 is to calculate the data variance var of each band image. j , including the following steps:

[0130] 2.1) Calculate the mean u of each feature band image j :where j represents the jth band, and the input hyperspectral image data X∈R m×n×l , where m and n represent the length and width of each band image, and l represents the number of bands in the hyperspectral image. For the data corresponding to the jth band feature, use x ab Represents the two-dimensional image data of this band, a∈[0,m], b∈[0,n], then the mean value u of the jth feature band image j It can be expressed as:

[0131]

[0132] 2.2) Calculate the data variance var of each characteristic band image j:Input hyperspectral image data X∈R m×n×l , the mean value u of each band image j , then the data variance of each characteristic band image can be expressed as:

[0133]

[0134] Where dist represents the band data x ab and the band mean u j The distance between them, Card(X) is the matrix basis of the hyperspectral image.

[0135] The present invention calculates the variance of each band and sorts the variance from small to large. Since the smaller the variance of the band data, the more conducive it is to abnormal target detection, the first N data are selected for abnormal target detection. The axis-parallel subspace selection method of the present invention greatly reduces data redundancy and reduces the computational complexity of detection.

[0136] Example 3

[0137] The method for detecting abnormal targets in a fast and robust hyperspectral image of the present invention is the same as that in Embodiment 1-2, except that the Mahalanobis distance D corresponding to the N groups of characteristic band data is calculated in step S2. Ma_i , including the following steps:

[0138] 3.1) Calculate the covariance matrix Γ of the i-th group of characteristic band data Xsub_i :Input the characteristic band data X of the i-th group sub_i ∈R m×n×q / N , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i_ 2D∈R q / N×mn , the covariance matrix Γ of the i-th group of characteristic band data Xsub_i ∈R q / N×q / N It can be expressed as:

[0139] Γ Xsub_i =X sub_i _2D×X sub_i _2D T (6)

[0140] 3.2) Calculate the mean matrix u of the i-th group of characteristic band data X_i :Input the characteristic band data X of the i-th group sub_i ∈R m×n×q / N , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i _2D∈R q / N×mn , the mean matrix u of the i-th group of characteristic band data X_i ∈R q / N×1 It can be expressed as:

[0141]

[0142] 3.3) Calculate the Mahalanobis distance D corresponding to N sets of feature band data Ma_i : Input the covariance matrix Γ of the i-th group of feature band data Xsub_i ∈R q / N×q / N , input the mean matrix u of the i-th group of feature band data X ∈R q / N×1 , the Mahalanobis distance D corresponding to the characteristic band data of the i-th group Ma_i_1 ∈R mn×1 :

[0143] Where a=1,2,…mn (8)

[0144] D Ma_i_1 The Mahalanobis distance graph D corresponding to the final i-th group of feature data is obtained by converting the mn×1-dimensional matrix into an m×n two-dimensional matrix. Ma_i ∈R m×n .

[0145] The present invention calculates the Mahalanobis distance anomaly map and uses the spectral anomaly characteristics of the abnormal target to preliminarily determine the position of the abnormal target. The characteristic data of each group can be used to calculate the Mahalanobis distance anomaly map in parallel, which greatly improves the detection time.

[0146] Example 4

[0147] The method for detecting abnormal targets in a fast and robust hyperspectral image of the present invention is the same as that in Embodiments 1-3, and the abnormal score map S based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data in step S3 is calculated. hist_i The steps include:

[0148] 4.1) Calculate the frequency histogram hist(k) of N groups based on the Mahalanobis distance map, and input the Mahalanobis distance map D of N groups of data Ma_1 , D Ma_2 , …D Ma_N , input the number of groups k of frequency histogram, each group contains the amount of data N_fix, D Ma_i ∈R m×n Convert to a one-dimensional array D Ma_i_1 ∈R mn×1 . Transform the one-dimensional array D Ma_i_1 ∈R mn×1 Sort from small to large and divide them into k groups. The amount of data in each group is N_fix. The maximum and minimum values ​​of each group of data are recorded as k_max and k_min respectively. Then the frequency histogram hist(k) of the kth group can be expressed as:

[0149]

[0150] 4.2) Calculate N groups of anomaly score graphs S based on frequency histograms hist_i : Input the hist(k) obtained by the Mahalanobis distance map of the i-th group, then the abnormal score result S based on the frequency histogram corresponding to the pixel points of the k-th group hist_i It can be expressed as:

[0151]

[0152] S hist_i_1 The final abnormal score result S based on the frequency histogram corresponding to the i-th group of feature data is obtained by converting the mn×1-dimensional matrix into an m×n two-dimensional matrix. hist_i ∈R m×n .

[0153] Based on the Mahalanobis distance anomaly score map, the present invention further obtains an anomaly score map based on a frequency histogram by utilizing the spatial anomaly characteristics of the abnormal target. The computational complexity of this step is very low, and the computational resources occupied are small, but the accuracy of anomaly detection can be greatly improved.

[0154] Example 5

[0155] The method for detecting abnormal targets in a fast and robust hyperspectral image of the present invention is the same as that in Embodiments 1-4, except that the abnormal score fusion result S of the Mahalanobis distance graph and the frequency histogram of each grouped data in step S4 is obtained. total_i The solution includes the following steps:

[0156] 5.1) Input the i-th group of anomaly score results S based on frequency histogram hist_i ∈R m×n , input the Mahalanobis distance graph D corresponding to the i-th group of feature data Ma_i ∈R m×n , then the anomaly score fusion result S of the i-th group total_i It is expressed as:

[0157]

[0158] In formula (11), S and λ D is the fusion coefficient.

[0159] 5.2) Determine the fusion coefficient λ S and λ D Value of: λ S and λ D They are respectively the matrix S based on the abnormal score map of the frequency histogram hist and the Mahalanobis distance anomaly map matrix D Ma The maximum eigenvalue of .

[0160] The present invention fully utilizes the spectral anomaly characteristics and spatial anomaly characteristics of the abnormal target by fusing the anomaly score map based on the Mahalanobis distance and the anomaly score map based on the frequency histogram, thereby improving the detection accuracy.

[0161] Example 6

[0162] The method for detecting abnormal targets in a fast and robust hyperspectral image of the present invention is the same as that in Embodiments 1-5, except that the binary accumulation B(a, b)∈R of the N sets of data is calculated in step S5. m×n The results include the following steps:

[0163] 6.1) Input the anomaly score fusion result S of N groups of data total_1 , S total_2 ,…S total_N , input threshold value Val threshold , solve the binary calculation result diagram M of the i-th group total_i (a, b), where a∈[0, m], b∈[0, n], i represents the i-th set of characteristic band data. total_i The calculation process is as follows:

[0164]

[0165] 6.2) Binary calculation result diagram M of input N sets of data total_1 , M total_2 , …M total_N The final binary accumulation of N sets of data B(a, b)∈R m×n , where a∈[0,m], b∈[0,n], the result can be expressed as:

[0166]

[0167] The present invention uses binary accumulation to assume that the positions of real abnormal targets in N groups of abnormal detection result graphs are fixed, while the abnormal targets detected by virtual detection appear randomly. Based on this assumption, as N groups of data gradually accumulate, the energy of the real abnormal target points will be accumulated each time, while the energy of the virtual detection targets is accumulated randomly. After binary accumulation, the energy difference between the real abnormal target and the virtual detection target position is large, and the energy peak in the binary accumulation result B (a, b) is searched as the abnormal detection result.

[0168] A more detailed example is given below to further illustrate the present invention:

[0169] Example 7

[0170] The fast and robust abnormal target detection method of hyperspectral images of the present invention is the same as that of Examples 1-6, with reference to Figure 1-4 The fast and robust abnormal target detection method of hyperspectral images of the present invention comprises the following steps:

[0171] Step 1: Use the axis-parallel subspace method to select the characteristic bands of the hyperspectral image: Input the hyperspectral image X∈R 100×100×205 , where the length and width of the two-dimensional image of each band are 100×100, and the number of bands of the input hyperspectral image is 205.

[0172] 1.1) Calculate the mean u of each feature band image j :where j represents the jth band, using x ab Represents the two-dimensional image data of this band, a∈[0,100], b∈[0,100], then the mean value u of the jth feature band image is j It can be expressed as:

[0173]

[0174] 1.2) Calculate the data variance var of each characteristic band image j :Input hyperspectral image data X∈R m×n×l , the mean value u of each band image j , then the data variance of each characteristic band image can be expressed as:

[0175]

[0176] Sorting by variance, 40 feature bands are selected starting from the band with the smallest variance as the subspace for abnormal target detection. The selected 40 subspaces are divided into 20 groups, each containing 2 subbands.

[0177] Step 2: Calculate the Mahalanobis distance D of each set of data Ma_i : Based on the 20 sets of data selected in step 1, step 2 is as follows:

[0178] 2.1) Calculate the covariance matrix Γ of the i-th group of characteristic band data Xsub_i :Input the characteristic band data X of the i-th group sub_i ∈R 100×100×2 , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i_ 2D∈R 2×10000 , the covariance matrix Γ of the i-th group of characteristic band data Xsub_i ∈R 2×2 It can be expressed as:

[0179] Γ Xsub_i =X sub_i _2D×X sub_i _2D T (16)

[0180] 2.2) Calculate the mean matrix u of the i-th group of characteristic band dataX_i :Input the characteristic band data X of the i-th group sub_i ∈R 100×100×2 , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i _2D∈R 2×10000 , that is, the data of each sub-band in the two sub-bands are arranged into a one-dimensional array with an array size of 1×10000. Then the mean matrix u of the i-th group of feature band data is X_i ∈R 2×1 It can be expressed as:

[0181]

[0182] Among them, a represents the number of bands, b represents the number of pixels in each band, a∈[0, 2], b∈[0, 10000].

[0183] 2.3) Calculate the Mahalanobis distance D corresponding to 20 sets of feature band data Ma_i : Input the covariance matrix Γ of the i-th group of feature band data Xsub_i ∈R 2×2 , input the mean matrix u of the i-th group of feature band data X ∈R 2×1 , the Mahalanobis distance D corresponding to the characteristic band data of the i-th group Ma_i_1 ∈R 10000×1 :

[0184] Where a=1,2,…10000(18)

[0185] D Ma_i_1 The Mahalanobis distance graph D corresponding to the final i-th group of feature data is obtained by converting the 10000×1-dimensional matrix into a 100×100 two-dimensional matrix. Ma_i ∈R m×n .

[0186] Step 3: Calculate the corresponding frequency histogram-based anomaly score graph S according to the 20 sets of Mahalanobis distance-based anomaly score graphs obtained. hist_i Step 3 is as follows:

[0187] 3.1) Input the 20 sets of Mahalanobis distance graphs D obtained in step 2 Ma_1 , D Ma_2 , …D Ma_20 , input the number of groups of frequency histogram 2000, each group contains data volume N_fix=5, D Ma_i ∈R 100×100 Convert to a one-dimensional array D Ma_i_1 ∈R 10000×1 . Transform the one-dimensional array D Ma_i_1 ∈R 100×1Sort from small to large and divide them into 2000 groups, with the data volume of each group being 5. The maximum and minimum values ​​of each group of data are recorded as k_max and k_min respectively. Then the frequency histogram hist(k) of the kth group can be expressed as:

[0188]

[0189] Where k is the number of groups of the frequency histogram, which is 2000.

[0190] 3.2) Calculate 20 sets of anomaly score graphs S based on frequency histograms hist_i : Input the hist(k) obtained from the Mahalanobis distance map of the i-th group, then the abnormal scores based on the frequency histogram corresponding to the 5 pixels in the k-th group are the same, which can be expressed as:

[0191]

[0192] S hist_i_1 The 10000×1-dimensional matrix is ​​converted into a 100×100 two-dimensional matrix to obtain the abnormal score result S based on the frequency histogram corresponding to the final i-th group of feature data hist_i ∈R m×n In S hist_i The larger the value corresponding to the pixel point, the more likely it is an abnormal target.

[0193] Step 4: Fusion of the abnormal score results of the Mahalanobis distance graph and frequency histogram of each grouped data: Based on the D obtained in steps 2 and 3 Ma_i and S hist_i , step 4 is as follows:

[0194] Input the i-th group of anomaly score results S based on frequency histogram hist_i ∈R 100×100 , input the Mahalanobis distance graph D corresponding to the i-th group of feature data Ma_i ∈R 100×100 , then the anomaly score fusion result S of the i-th group total_i It is expressed as:

[0195]

[0196] where λ S =4 and λ D =1 is the fusion coefficient, and are the maximum eigenvalues ​​of the frequency histogram score matrix and the Mahalanobis distance map score matrix respectively.

[0197] Step 5: Calculate the binary accumulation of 20 sets of data B(a, b)∈R 100×100 Results. According to the 20 groups of fused anomaly detection result graphs obtained in step 4, the specific calculation steps are as follows:

[0198] 5.1) Input the anomaly score fusion result S of 20 sets of data total_1 , S total_2 ,…S total_20 , input threshold value Val threshold = 0.5, solve the binary calculation result of group i M total_i (a, b), where (a, b) represents the pixel coordinate position, a∈[0, 100], b∈[0, 100], i represents the i-th group of features, i∈[0, 20]. M total_i The calculation process is as follows:

[0199]

[0200] 5.2) Binary calculation result diagram M of input N sets of data total_1 , M total_2 , …M total_N The final binary accumulation of N sets of data B(a, b)∈R 100×100 , where a∈[0,100], b∈[0,100], N=20, the result can be expressed as:

[0201]

[0202] The higher the energy value in the binary accumulation result B(a, b), the more likely it is that it is an abnormal target that is finally detected.

[0203] In the present invention, the abnormal target detection method of the hyperspectral image not only uses the spectral abnormal information of the abnormal target but also combines the spatial abnormal information of the abnormal target, greatly improving the detection accuracy by increasing the computational complexity very little. Secondly, the present invention uses the method of subspace grouping and binary accumulation, so that each group of abnormal detection is parallelized, reducing the detection time and increasing the robustness of the abnormal detection algorithm. The hyperspectral abnormal target detection algorithm proposed in the present invention has great engineering practical significance.

[0204] The technical effects of the present invention are described below through simulation.

[0205] Example 8

[0206] The fast and robust abnormal target detection method of hyperspectral images of the present invention is the same as that of Examples 1-7.

[0207] The technical effects of the present invention are described in detail below in conjunction with simulation experiments.

[0208] 1. Simulation conditions:

[0209] The detection results of each step of the input of the same hyperspectral image data are analyzed.

[0210] Given a hyperspectral image X∈R 100×100×205, the number of selected bands is 40, each 2 bands is a group, a total of 20 groups, the threshold Val threshold =0.5,λ S , D are the fusion coefficients of the anomaly map based on Mahalanobis distance and the anomaly map based on frequency histogram, which are 4 and 1 respectively.

[0211] 2. Simulation results and analysis:

[0212] See also Figure 6 , Figure 6 These are 20 groups of abnormal score graphs based on Mahalanobis distance obtained by the present invention. Since the background of the hyperspectral image does not necessarily obey the Gaussian distribution, it can be seen from the figure that the abnormal targets detected using only the spectral abnormal information of the abnormal targets are different from those detected using the spectral abnormal information of the abnormal targets. Figure 5 The effect of anomaly detection is very poor compared with the actual abnormal target location.

[0213] See also Figure 7 , Figure 7 The present invention obtains 20 groups of abnormal score graphs based on frequency histograms. The bright area in the abnormal score graph of each group of data is the position of the detected abnormal target. It can be seen that the abnormal score graph has been significantly improved after using the spatial abnormal information of the abnormal target. However, there are false detection targets in each group of data in the abnormal score graph, and the positions of the false detection targets are different.

[0214] See also Figure 8 , Figure 8 It is the result of fusing the abnormal score map based on Mahalanobis distance and the abnormal score map based on frequency histogram. Figure 6 The anomaly score map based on Mahalanobis distance is very poor, so the fused result is equivalent to the anomaly score map based on frequency histogram.

[0215] See also Fig. 9 , Fig. 9 is the abnormal score map after fusion of 20 groups, at the threshold Val threshold =0.5, where the position of 1 represents the abnormal target and the position of 0 represents the background.

[0216] See also Fig.10 , Fig.10 This is the result of binary accumulation of 20 sets of data. It can be seen from the figure that the anomaly score map after binary accumulation is already quite close to the position of the actual anomaly target.

[0217] The present invention provides a fast and robust method for detecting abnormal targets in hyperspectral images. By combining the spectral anomaly information and spatial anomaly information of the abnormal targets, the accuracy of anomaly detection is improved. The detection speed and the robustness of the enhanced detection algorithm are improved based on subspace grouping and binary accumulation, respectively. The present invention provides new prospects for the industrial application of hyperspectral images due to computing power limitations.

[0218] In short, the invention discloses a fast and robust abnormal target detection method for hyperspectral images, which mainly solves the defects that the detection accuracy and detection complexity cannot be taken into account at the same time and the detection method has poor robustness due to the large amount of hyperspectral image data and the lack of target prior information, and realizes a fast, accurate and robust abnormal detection method. The implementation process is: using the axis-parallel subspace method to select the characteristic bands of the hyperspectral image; the selected q characteristic bands are divided into N groups, and the Mahalanobis distance is calculated for each group of data; the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data is solved; the abnormal score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused; the binary accumulation results of the N groups of data are calculated, and abnormal target detection is realized through the energy threshold. The present invention can be applied to the rapid processing of abnormal target detection in hyperspectral images, which makes up for the shortcomings of the existing high-precision abnormal target detection methods with high computational complexity and slow detection speed, and lays a foundation for the identification of the wide application of real-time processing of hyperspectral images.

[0219] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0220] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A fast and robust method for abnormal target detection in hyperspectral images, characterized in that: The fast and robust hyperspectral image abnormal target detection method comprises the following steps: Step 1: Use the axis-parallel subspace method to select the characteristic bands of the hyperspectral image; Step 2: divide the selected q feature bands into N groups, and calculate the Mahalanobis distance for each group of data; Step 3, solving the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data; Step 4: Fusion the abnormality scores of the Mahalanobis distance graph and the frequency histogram of each grouped data; Step 5: Calculate the binary accumulation results of N groups of data and realize abnormal target detection through energy threshold; In step 3, the abnormal score graph S based on the frequency histogram corresponding to the Mahalanobis distance graph of the N groups of data is solved hist_i ,include: Input the Mahalanobis distance graph D of group i Ma_i , the number of groups k in the frequency histogram; convert the data in the Mahalanobis distance map into a one-dimensional array and sort D Mai _1d, fix the amount of data contained in each group to N_fix, calculate the starting value and cutoff value of each group of data in the frequency histogram, and obtain the width of each group. Then the height of each group in the frequency histogram histogram (k) can be obtained by N_fix and the width of each group. Finally, the abnormal score of each group based on the frequency histogram is S hist_i ,include: Input the Mahalanobis distance graph D of N sets of data Ma_1 , D Ma_2 , …D Ma_N , input the hist(k) obtained by the Mahalanobis distance map of the i-th group, then the abnormal score result S based on the frequency histogram corresponding to the pixel points of the k-th group hist_i It is expressed as: S hist_i_1 The final abnormal score result S based on the frequency histogram corresponding to the i-th group of feature data is obtained by converting the mn×1-dimensional matrix into an m×n two-dimensional matrix. hist_i ∈R m×n .

2. The fast and robust hyperspectral image abnormal target detection method according to claim 1, characterized in that: In step 1, the method of using axis-parallel subspace to select characteristic bands of the hyperspectral image includes: The hyperspectral image is represented as X∈R m×n×l ; where m and n represent the length and width of each band image, and l represents the number of bands in the hyperspectral image; the value of the abnormal target is greatly different from the background value and is sparsely distributed, so the more concentrated the background pixel value is, the easier it is to detect the abnormal target away from the background. Based on this cognition, the data variance var of each characteristic band image is calculated j , sort by variance, and select q feature bands starting from the band with the smallest variance as the subspace for abnormal target detection; Among them, the data variance var of each characteristic band image j Calculations include: (1) Calculate the mean u of each feature band image j : j represents the jth characteristic band, x ab Represents the pixel value of the ath row and bth column of the jth feature band, then the mean value u of each feature band image is determined j for: (2) Calculate the data variance var of each characteristic band image j :Where dist represents the band data and the mean u of the band j The distance between them, Card(X) is the matrix basis of the hyperspectral image, and the data variance of each characteristic band image is expressed as:

3. The fast and robust hyperspectral image abnormal target detection method according to claim 1, characterized in that: In step 2, the selected q characteristic bands are evenly divided into N groups, and the Mahalanobis distance is calculated for each group of data, including: Input the selected q characteristic band data X sub ∈R m×n×q , the data is evenly divided into N groups, the i-th group is represented by: X sub_i ∈R m×n×q / N ; Calculate the Mahalanobis distance D of the i-th group of subspace feature data Ma_i ; Wherein, the Mahalanobis distance D corresponding to the N groups of characteristic band data is calculated Ma_i ,include: (1) Calculate the covariance matrix Γ of the i-th group of characteristic band data Xsub_i :Input the characteristic band data X of the i-th group sub_i ∈R m×n×q / N , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i _2D∈R q / N×mn , the covariance matrix Γ of the i-th group of characteristic band data Xsub_i ∈R q / N×q / N It is expressed as: Γ Xsub_i =X sub_i _2D×X sub_i _2D Τ ; (2) Calculate the mean matrix u of the i-th group of characteristic band data X_i :Input the characteristic band data X of the i-th group sub_i ∈R m ×n×q / N , the three-dimensional X sub_i Convert to a two-dimensional matrix X sub_i _2D∈R q / N×mn , the mean matrix u of the i-th group of characteristic band data X_i ∈R q / N×1 It is expressed as: (3) Calculate the Mahalanobis distance D corresponding to N sets of feature band data Ma_i : Input the covariance matrix Γ of the i-th group of feature band data Xsub_i ∈R q / N×q / N , input the mean matrix u of the i-th group of feature band data X ∈R q / N×1 , the Mahalanobis distance D corresponding to the characteristic band data of the i-th group Ma_i_1 ∈R mn×1 : Where a=1, 2, ... mn; D Ma_i_1 The Mahalanobis distance graph D corresponding to the final i-th group of feature data is obtained by converting the mn×1-dimensional matrix into an m×n two-dimensional matrix. Ma_i ∈R m×n .

4. The fast and robust hyperspectral image abnormal target detection method according to claim 1, characterized in that: In step 4, the anomaly score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused, including inputting the Mahalanobis distance map D Ma_i and the anomaly score map S based on the frequency histogram hist_i , the fusion result is S total_i ; The abnormal score fusion result S of the Mahalanobis distance graph and the frequency histogram of each grouped data is total_i Solving, including: Input the i-th group of anomaly score results S based on frequency histogram hist_i ∈R m×n , input the Mahalanobis distance graph D corresponding to the i-th group of feature data Ma_i ∈R m×n , then the anomaly score fusion result S of the i-th group total_i It is expressed as: Among them, λ S and λ D is the fusion coefficient.

5. The fast and robust hyperspectral image abnormal target detection method according to claim 1, characterized in that: In step 5, the binary accumulation results of N groups of data are calculated, and abnormal target detection is achieved through energy threshold, including: (1) Calculate the binary accumulation of N sets of data B(a, b)∈R m×n Results: Fusion anomaly score map S of input N groups of data total_i , set the threshold value Val threshold , greater than Val threshold The pixel position is considered to be the location of the abnormal target, which is less than Val threshold The pixel position of each group is considered as the background, and the abnormal target pixel position of each group is assigned 1, and the other background pixels are assigned 0, and the abnormal detection results of N groups of data are accumulated; (2) Determine the anomaly detection target: Input the binary accumulation result of N groups of data, design the energy threshold, and consider the pixels with energy greater than the threshold as anomaly targets, and the pixels with energy less than the threshold as background. Wherein, the binary accumulation B(a, b)∈R of the N sets of data is calculated m×n Results, including: 1) Input the anomaly score fusion result S of N groups of data total_1 , S total_2 ,…S total_N , input threshold value Val threshold , the binary calculation result diagram of the i-th group is M total_i (a, b); (a, b) represents the coordinate position of the pixel, i represents the i-th group of characteristic band data; M total_i The calculation process is as follows: 2) Input N sets of data and the binary calculation result graph M total_1 , M total_2 , …M total_N , the final binary accumulation of N groups of data B(a, b)∈R m×n The results are expressed as:

6. A fast and robust hyperspectral image abnormal target detection system implementing the fast and robust hyperspectral image abnormal target detection method according to any one of claims 1 to 5, characterized in that: The fast and robust hyperspectral image abnormal target detection system comprises: The feature band selection module is used to select the feature bands of the hyperspectral image using the axis-parallel subspace method; A Mahalanobis distance calculation module is used to divide the selected q characteristic bands into N groups and calculate the Mahalanobis distance for each group of data; An abnormal score map acquisition module is used to solve the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of N groups of data; The anomaly score fusion module is used to fuse the anomaly scores of the Mahalanobis distance map and frequency histogram of each grouped data; The abnormal target detection module is used to calculate the binary accumulation results of N groups of data and realize abnormal target detection through energy threshold.

7. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the fast and robust hyperspectral image abnormal target detection method according to any one of claims 1 to 5, comprising the following steps: The axis-parallel subspace method is used to select the characteristic bands of the hyperspectral image; the selected q characteristic bands are divided into N groups, and the Mahalanobis distance is calculated for each group of data; the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data is solved; the abnormal score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused; Calculate the binary accumulation results of N groups of data and realize abnormal target detection through energy threshold.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the fast and robust hyperspectral image abnormal target detection method according to any one of claims 1 to 5, comprising the following steps: The axis-parallel subspace method is used to select the characteristic bands of the hyperspectral image; the selected q characteristic bands are divided into N groups, and the Mahalanobis distance is calculated for each group of data; the abnormal score map based on the frequency histogram corresponding to the Mahalanobis distance map of the N groups of data is solved; the abnormal score of the Mahalanobis distance map and the frequency histogram of each grouped data is fused; the binary accumulation results of N groups of data are calculated, and abnormal target detection is achieved through energy threshold.

9. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the fast and robust hyperspectral image abnormal target detection system as described in claim 6.

Images