A Hyperspectral Anomaly Detection Method Based on Joint Spatial-Spectral Salient Feature Representation

Abstract

The present invention relates to an anomaly detection method based on hyperspectral images. The present invention performs anomaly detection on hyperspectral images based on a method of expressing joint spatial-spectral saliency features. The specific method includes: First, the extraction of spectral features is completed through a spectral feature extraction method that combines a dual-window structure and an improved spectral distance calculation. At the same time, the extraction of spatial features is completed through a spatial feature extraction method based on data dimensionality reduction and curvature filtering. Secondly, the joint spatial-spectral initial features are obtained by combining two different attribute features. Finally, anomaly detection is performed through a hyperspectral saliency detection method based on context awareness, thereby obtaining the anomaly detection result.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing technology, and particularly relates to a hyperspectral anomaly detection method based on an air-spectrum joint saliency feature expression. Background Technique

[0002] Hyperspectral imaging has the advantages of high spatial resolution and wide spectral range. The spectral detection range covers the visible light to infrared light bands, and has the characteristics of a large number of approximate continuous spectral information with a wide spectral range. At the same time, it also contains relatively detailed spatial information of ground objects. Therefore, its spatial information and spectral information can be fully utilized to complete the functions of detection and analysis. At the same time, with the continuous development of optical imaging technology, the spectral resolution ability and spatial resolution ability of hyperspectral imaging spectrometers are also continuously improving, and the application fields of hyperspectral remote sensing are also continuously expanding.

[0003] There are a large number of characteristic information with different attributes in hyperspectral images, and the most widely used among them include spectral features and spatial features. When using spectral features, although spectral information can reflect the true attributes of targets, the phenomena of "same object with different spectra" and "same spectrum with different objects" will lead to detection errors; when detecting through the spatial distribution information of ground objects in hyperspectral images, other components such as background and noise will inevitably be left. In addition, due to the fact that the shape and distribution of objects cannot fully reflect the true attributes of objects, the existence of camouflaged targets will bring false alarm phenomena. Summary of the Invention

[0004] In order to overcome the above-mentioned shortcomings of the prior art, the main object of the present invention is to provide a hyperspectral anomaly detection method based on an air-spectrum joint saliency feature expression, in order to solve the problems such as large errors and low accuracy existing in the detection of the spatial distribution information of ground objects in hyperspectral images by anomaly detection algorithms based on single features.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A hyperspectral image anomaly detection method based on an air-spectrum joint saliency feature expression, comprising the following steps:

[0007] Step 1, perform background estimation on the original hyperspectral image using a double-window structure, and calculate the spectral distance between the local background and the pixel to be measured;

[0008] Step 2, calculate the relative spectral difference between each pixel corresponding spectral dimension and the surrounding background by sliding the double-window structure to obtain an initial spectral feature map;

[0009] Step 3: By using the principal component analysis method, redundant information in the original hyperspectral image is removed to obtain the reduced-dimension images of multiple bands. Then, the reduced-dimension images of multiple bands are weighted and fused to obtain the fused reduced-dimension image;

[0010] Step 4: For the fused reduced-dimension image, total variation curvature filtering is used for energy minimization calculation to obtain the initial hyperspectral spatial feature map, that is, the two-dimensional ground object distribution feature map;

[0011] Step 5: The initial spectral feature map obtained in Step 2 is fused with the initial hyperspectral spatial feature map obtained in Step 4 to obtain the initial hyperspectral spatial-spectral joint feature map;

[0012] Step 6: The method for detecting hyperspectral saliency anomaly features based on context awareness is used to detect the targets in the initial hyperspectral spatial-spectral joint feature map obtained in Step 5, and the detection results are obtained.

[0013] The present invention adopts a spectral feature extraction method combining a joint double-window structure and improved spectral distance calculation and a spatial feature extraction method based on data dimensionality reduction and curvature filtering to respectively extract and obtain spectral features and spatial features. Then, the spectral features and spatial features are dot-multiplied and fused to obtain the initial spatial-spectral joint features. Finally, considering the saliency characteristics of abnormal targets, a hyperspectral saliency feature detection algorithm is studied and implemented, and subsequent detection is completed based on the initial spatial-spectral joint features. Through comparative simulation experiments with other classical anomaly detection algorithms on different datasets, the simulation results show that the comprehensive performance of the present invention on different datasets is better than that of the comparative algorithms, indicating that the present invention has good anomaly detection capabilities. Description of the Drawings

[0014] Figure 1 It is a flow chart of the present invention.

[0015] Figure 2 It is a schematic diagram of a double-window structure.

[0016] Figure 3 It is a schematic diagram of hyperspectral image preprocessing.

[0017] Figure 4 It is a flow chart of joint feature extraction.

[0018] Figure 5 It is an example diagram of the present invention. Among them, (a) is the pseudo-color map of the original image, and (b) is the abnormal distribution map.

[0019] Figure 6 It is a diagram of the anomaly detection result in the example of the present invention. Among them, (a) is the pseudo-color map of the RPCA-RX detection result, and (b) is the pseudo-color map of the detection result of the present invention.

[0020] Figure 7This is the three-dimensional curve graph of the anomaly detection results in the embodiments of the present invention. Among them, (a) is the three-dimensional kurtosis graph of the RPCA-RX detection results, and (b) is the three-dimensional kurtosis graph of the detection results of the present invention. Detailed implementation manners

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and 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.

[0022] As described above, when using spectral features, "same object with different spectra" and "same spectra with different objects" will inevitably lead to detection errors. Especially when detecting the ground object spatial distribution information of hyperspectral images, a large number of camouflage targets may be generated, greatly reducing the accuracy of the detection results. Obviously, there are many such defects in the anomaly detection algorithms based on single features, such as being easily affected by external factors and insufficient utilization of the significant characteristics of anomaly targets in traditional algorithms. Through analysis, it can be seen that the features of anomaly targets have certain "significant characteristics". Therefore, the present invention proposes an anomaly detection algorithm based on the expression of joint spatio-spectral significant features. First, a spectral feature extraction method using a joint double-window structure and improved spectral distance calculation and a spatial feature extraction method based on data dimensionality reduction and curvature filtering are adopted to respectively extract spectral features and spatial features, and then the spectral features and spatial features are dot-multiplied and fused to obtain the initial joint spatio-spectral features. Finally, considering the significant characteristics of anomaly targets, an improved hyperspectral significant feature detection algorithm is proposed, and subsequent detection is completed based on the initial joint spatio-spectral features. Through comparative simulation experiments with other classical anomaly detection algorithms, the simulation results show that the comprehensive performance of this method is better than that of the comparative algorithms, indicating that this method has good anomaly detection capabilities.

[0023] To demonstrate the effectiveness of the method of the present invention, real hyperspectral images are used for anomaly detection. The experiment uses a sub-image with a size of 100×100 and 189 bands in the San Diego airport scene image collected by an airborne visible / infrared imaging spectrometer (AVIRIS) sensor for anomaly detection, and three airplanes in this sub-image are used as anomaly targets. Refer to Figure 1 , the basic process of anomaly detection for the SanDiego_Plane image data is as follows:

[0024] Step 1, the background of the original hyperspectral image is estimated using a double-window structure, and the spectral distance between the local background and the pixel to be measured is calculated.

[0025] In the present invention, the original hyperspectral image refers to the SanDiego_Plane image, and the local background refers to Figure 2The area between the inner window and the outer window. The pixel to be measured is located in the inner window. Each pixel is calculated by sliding the window through traversal. During the calculation, this pixel is the corresponding pixel to be measured.

[0026] The specific steps of the above calculation are as follows:

[0027] Step 1.1, the original hyperspectral image Y ∈ R M×N×B where the spatial dimension size is M×N, M represents the number of rows contained in the hyperspectral image, N represents the number of columns contained in the hyperspectral image, and B is the number of bands. Each pixel in the space corresponds to a spectral curve of 1×1×B;

[0028] Step 1.2, for the pixel at coordinates (i,j) in the hyperspectral image Y ∈ R M×N×B where the corresponding spectral curve is l, a double-window structure is introduced with this point as the center point. Refer to Figure 2 , the inner window range is the protection area. The inner window is placed in the outer window, and the area between the outer window and the inner window is the background area. The pixel to be measured is located in the inner window. Exemplarily, the sizes of the inner and outer windows w in 、w out can be 9 and 11 respectively. The average spectral curve l mean of the pixels between the inner and outer windows is obtained, and the spectral distance d(X,Y) between l and l mean (i.e., the spectral distance between the local background and the pixel to be measured) is calculated according to the following formula as the initial spectral feature value corresponding to the pixel at (i,j);

[0029]

[0030] In the formula, X and Y are two different vectors, n is the number of dimensions contained in the vector, the one-dimensional vector X = [x1,x2,x3,…,x k ,…,x n , Y = [y1,y2,y3,…,y k ,…,y n , num is the adaptive weight coefficient, and the value of num is determined according to the following formula:

[0031]

[0032] Step 2, calculate the relative spectral difference between the spectral dimension corresponding to each pixel and the surrounding background by sliding the double-window structure. After calculating each pixel in turn, a two-dimensional feature matrix can be obtained, that is, the initial spectral dimension feature map R spectrum ∈R M×N .

[0033] Step 3, through principal component analysis, redundant information in the original hyperspectral image is removed to obtain the reduced-dimension images of multiple bands, and then the reduced-dimension images of multiple bands are weighted and fused to obtain the fused reduced-dimension image. Refer to Figure 3 , which specifically includes the following steps;

[0034] Step 3.1, use the PCL dimensionality reduction algorithm to perform dimensionality reduction on the original hyperspectral image. After dimensionality reduction, the data dimension is m, and the m eigenvectors correspond to m maximum eigenvalues, which are {λ1, λ 2, …, λ m}, and the corresponding eigenvectors {W1, W2, …, W m} in the eigenmatrix form a new projection matrix.

[0035] In the present invention, m is selected according to the information contribution of the eigenvector corresponding to the eigenvalue. After the cumulative information contribution rate of the current m maximum eigenvalues corresponding to the principal components reaches the set threshold, these m principal component data are retained, thus completing the data dimensionality reduction. Exemplarily, the threshold can be selected as 97%. That is, when the sum of the current m eigenvalues is greater than 97%, the principal component data corresponding to the subsequent lower eigenvalues are no longer retained, and the principal component data with the cumulative information contribution rate of the first m reaching more than 97% are retained.

[0036] Step 3.2, use the projection matrix obtained in Step 3.1 to calculate the corresponding m two-dimensional images with the original high-dimensional data, that is, the first m principal components are {Y1, Y2, …, Y m}, that is, the reduced-dimension images of m bands, which contain most of the features in the original hyperspectral data.

[0037] Step 3.3, according to the following formula, for the reduced-dimension images of m bands, weighted fusion is performed according to the corresponding eigenvalue size to obtain the fused reduced-dimension image, that is, the initial characteristic map Y of the representative two-dimensional ground object distribution final .

[0038]

[0039] Step 4, use total variation curvature filtering to operate on the reduced-dimension hyperspectral image obtained in Step 3 to obtain the initial abnormal target two-dimensional spatial characteristic map R space ∈R M×N , refer to Figure 4 , which specifically includes the following steps,

[0040] Step 4.1, calculate the total variation curvature filtering to construct the projection operators d1~d8 according to the following formula:

[0041]

[0042] Wherein, O represents the two-dimensional matrix image to be processed; O(i,j) is the pixel information at the corresponding position coordinates (i,j) in the image; O(i+1,j-1) is the pixel information at the corresponding position coordinates (i+1,j-1) in the image; O(i+1,j) is the pixel information at the corresponding position coordinates (i+1,j) in the image; O(i+1,j+1) is the pixel information at the corresponding position coordinates (i+1,j+1) in the image; O(i,j-1) is the pixel information at the corresponding position coordinates (i,j-1) in the image; O(i,j+1) is the pixel information at the corresponding position coordinates (i,j+1) in the image; O(i-1,j-1) is the pixel information at the corresponding position coordinates (i-1,j-1) in the image; O(i-1,j) is the pixel information at the corresponding position coordinates (i-1,j) in the image; O(i-1,j+1) is the pixel information at the corresponding position coordinates (i-1,j+1) in the image; O(i,j+1) is the pixel information at the corresponding position coordinates (i,j+1) in the image; d1 to d8 are the projection distances, representing the distance between the central pixel of the 3×3 size sliding local box and the corresponding direction curvature plane;

[0043] Step 4.2, calculate the minimum projection distance d of the pixel variable at (i,j) according to the following formula min :

[0044] d min = min{|d1|, |d2|,..., |d8|}

[0045] Step 4.3, correct the central pixel of the local sliding box according to the following formula to obtain the corrected image as B(i,j):

[0046] B(i,j) = O(i,j) + d min

[0047] Step 4.4, perform a subtraction operation on the original image O(i,j) and the background image B(i,j) according to the following formula to suppress the background components in the original image and highlight the abnormal target distribution:

[0048] R space = R(i, j) = |O(i, j) - B(i, j)| 2

[0049] Wherein, R(i, j) and R space are the results after the subtraction operation of the original image and the background image, that is, the initial hyperspectral spatial feature map, which is used as the initial two-dimensional spatial feature map of the abnormal target in the hyperspectral image, R space ∈R M×N .

[0050] Step 5, refer to Figure 4, the initial spectral feature map R obtained in step 2 is calculated according to the following formula spectrum ∈R M×N is dot-multiplied and fused with the initial hyperspectral spatial feature map R space ∈R M×N to obtain the initial hyperspectral spatio-spectral joint feature map F ∈ R M×N ;

[0051] F = R space ·R spectrum

[0052] In the formula, F represents the initial hyperspectral spatio-spectral joint feature map, R spectrum represents the initial spectral feature map, R space represents the initial hyperspectral spatial feature map, · represents the dot-multiplication operation, that is, the elements at the same position of two matrices are multiplied to obtain the result of a new matrix.

[0053] In step 6, a method for detecting hyperspectral saliency anomaly features based on context awareness is used to detect the targets in the initial hyperspectral spatio-spectral joint feature map obtained in step 5, and the detection results are obtained.

[0054] The specific steps are as follows:

[0055] Step 6.1, calculate the saliency difference value d(F(i, j), F(i1, j1)) between different points at coordinates (i, j) and (i1, j1) in the initial hyperspectral spatio-spectral joint feature map F ∈ R M×N according to the following formula:

[0056]

[0057] In the formula, c represents the weight factor, and d represents the saliency difference value between two points. d position is the Euclidean distance between coordinates (i, j) and (i1, j1), and c = 10;

[0058] Step 6.2, according to the following formula, the sum of the saliency difference values between a point and the global image is used as the saliency value result S(i, j) at point (i, j), and the initial saliency value image S is obtained by traversing all the pixels of the image in turn;

[0059]

[0060] Step 6.3, according to the following formula, set the threshold th = 0.02, mark the elements in the initial saliency value image S that are greater than the threshold as significantly abnormal elements, and update the saliency value according to the Euclidean distance between each point and the nearest adjacent significantly abnormal element as the weight factor to obtain a new saliency image

[0061]

[0062] In the formula, is the new significant value after background suppression based on the context before and after, and d near is the Euclidean distance between the point (i,j) and the nearest significant abnormal element nearby;

[0063] Step 6.4, according to the new significant image judge the type of each element and output the detection result.

[0064] Figure 5 Fig. shows an example diagram of the present invention. (a) and (b) are respectively the pseudo-color diagram of the original image and the abnormal distribution diagram. Figure 6 In (a) and (b), they are respectively the abnormal detection results of the RPCA_RX method and the method of the present invention. It can be seen that the abnormal targets are not obvious in the detection result of the RPCA_RX method, and there are more background residues. The targets are obvious in the detection result of the method of the present invention; Figure 7 In (a) and (b), they are respectively the three-dimensional grid diagrams corresponding to the detection results of the RPCA_RX method and the method of the present invention. It can be seen that in the detection result of the RPCA_RX method, there is a numerical peak at the corresponding position of the abnormal target, but the target is not very prominent. The method of the present invention has a better background suppression effect and can better detect the abnormal target information in the hyperspectral image.

Claims

1. An anomaly detection method for hyperspectral images based on joint spatial-spectral saliency feature representation, characterized in that It includes the following steps: Step 1, perform background estimation on the original hyperspectral image using a double-window structure, and calculate the spectral distance between the local background and the pixel to be measured; Step 2, calculate the relative spectral difference between each pixel's corresponding spectral dimension and the surrounding background by sliding the double-window structure to obtain an initial spectral feature map; Step 3, through principal component analysis, remove the redundant information in the original hyperspectral image to obtain a reduced-dimension image of multiple bands, and then perform weighted fusion on the reduced-dimension images of multiple bands to obtain a fused reduced-dimension image; Step 4, perform energy minimization calculation on the fused reduced-dimension image using total variation curvature filtering to obtain an initial hyperspectral spatial feature map, that is, a two-dimensional ground object distribution feature map; Step 5, fuse the initial spectral feature map obtained in Step 2 with the initial hyperspectral spatial feature map obtained in Step 4 to obtain an initial hyperspectral spatial-spectral joint feature map; Step 6, use a context-aware hyperspectral saliency anomaly feature detection method to detect the target in the initial hyperspectral spatial-spectral joint feature map obtained in Step 5 to obtain a detection result.

2. The anomaly detection method for hyperspectral images based on the joint spatial-spectral saliency feature expression according to claim 1, characterized in that In step 1, for the pixel at coordinate (i, j) in the original hyperspectral image Y ∈ R M×N×B , with the corresponding spectral curve being l, a double-window structure is introduced with this point as the center point, and the average spectral curve l mean of the pixels between the inner and outer windows is obtained. Then, calculate the spectral distance between l and l mean according to the following formula. The spectral distance between the local background and the pixel to be measured, i.e., d(X, Y), is used as the initial spectral feature value corresponding to the pixel at (i, j): Among them, M represents the number of rows included in the hyperspectral image, N represents the number of columns included in the hyperspectral image, B is the number of bands, X and Y are two different vectors, n is the number of dimensions included in the vector, the one-dimensional vector X = [x1, x2, x3, …, x k , …, x n , Y = [y1, y2, y3, …, y k , …, y n , num is the adaptive weight coefficient, and the value of num is determined according to the following formula:

3. The anomaly detection method for hyperspectral images based on the expression of joint spatial-spectral saliency features according to claim 2, characterized in that, In step 2, after calculating each pixel in sequence, a two-dimensional feature matrix, namely the initial spectral feature map R, is obtained. spectrum ∈R M×N .

4. The anomaly detection method for hyperspectral images based on the expression of joint spatial-spectral saliency features according to claim 1, characterized in that, The said Step 3 includes: Step 3.1, use the PCL dimensionality reduction algorithm to perform dimensionality reduction on the original hyperspectral image. After dimensionality reduction, the data dimensionality is m, and the m eigenvectors correspond to m maximum eigenvalues, which are {λ1, λ2, …, λ m}, and the corresponding eigenvectors {W1, W2, …, W m} in the eigenmatrix form a new projection matrix; Step 3.2: Use the projection matrix obtained in Step 3.1 to calculate the corresponding m two-dimensional images from the original high-dimensional data, that is, the first m principal components {Y1, Y2, …, Y m}, which are also the images after dimensionality reduction of m bands and contain most of the features in the original hyperspectral image; Step 3.3, according to the following formula, perform weighted fusion on the dimension-reduced images of m bands according to the magnitudes of their corresponding eigenvalues to obtain a fused dimension-reduced image, that is, an initial characteristic map Y of the representative two-dimensional ground object distribution final : 。 5. The anomaly detection method for hyperspectral images based on the expression of joint spatial-spectral saliency features according to claim 4, characterized in that In Step 3.1, the dimension m of the reduced-dimension data is selected according to the information contribution of the eigenvector corresponding to the eigenvalue. After the cumulative information contribution rate of the current m largest eigenvalues corresponding to the principal components reaches a set threshold, these m principal component data are retained, thereby completing data reduction.

6. The anomaly detection method for hyperspectral images based on the expression of joint spatial-spectral saliency features according to claim 4, wherein The said threshold is 97%. In this algorithm, when the cumulative sum of the current m eigenvalues is greater than 97%, the principal component data corresponding to the subsequent lower eigenvalues are no longer retained, that is, the first m principal component data with a cumulative information contribution rate of more than 97% are retained.

7. The anomaly detection method for hyperspectral images based on the joint spatial-spectral saliency feature expression according to claim 4, wherein The said Step 4 includes: Step 4.1, calculate the total variation curvature filtering to construct projection operators d1~d8 according to the following formula: In the formula, O represents the two-dimensional matrix image to be processed; O(i,j) is the pixel information at the position with coordinates (i,j) in the image; O(i+1,j-1) is the pixel information at the position with coordinates (i+1,j-1) in the image; O(i+1,j) is the pixel information at the position with coordinates (i+1,j) in the image; O(i+1,j+1) is the pixel information at the position with coordinates (i+1,j+1) in the image; O(i,j-1) is the pixel information at the position with coordinates (i,j-1) in the image; O(i,j+1) is the pixel information at the position with coordinates (i,j+1) in the image; O(i-1,j-1) is the pixel information at the position with coordinates (i-1,j-1) in the image; O(i-1,j) is the pixel information at the position with coordinates (i-1,j) in the image; O(i-1,j+1) is the pixel information at the position with coordinates (i-1,j+1) in the image; O(i,j+1) is the pixel information at the position with coordinates (i,j+1) in the image; d1~d8 are the projection distances, representing the distance between the central pixel of the 3×3 size sliding local frame and the curvature plane in the corresponding direction; Step 4.2, calculate the minimum projection distance d of the pixel variable at (i, j) according to the following formula min ; d min = min{|d1|, |d2|,..., |d8|} Step 4.3, correct the center pixel of the local sliding window according to the following formula to obtain the corrected image as B(i, j); B(i,j) = O(i,j) + d min Step 4.4, perform a subtraction operation on the original image O(i, j) and the background image B(i, j) according to the following formula to suppress the background components in the original image and highlight the abnormal target distribution: R space = R(i, j) = |O(i, j) - B(i, j)| 2 Wherein, R(i,j) and R space are the results after subtracting the background image from the original image, that is, the initial hyperspectral spatial feature map, which is used as the initial two-dimensional spatial feature map of abnormal targets in the hyperspectral image. R space ∈R M×N .

8. The anomaly detection method for hyperspectral images based on the expression of joint spatial-spectral saliency features according to claim 1, characterized in that For the said Step 5, the fusion formula is as follows: F = R space ·R spectrum In the formula, F represents the initial hyperspectral spatial-spectral joint feature map, and R spectrum represents the initial spectral feature map, and R space represents the initial hyperspectral spatial feature map. · represents the dot product operation, that is, the elements at the same positions of two matrices are multiplied to obtain a new matrix result.

9. The anomaly detection method for hyperspectral images based on the expression of joint spatial-spectral saliency features according to claim 1, wherein The said Step 6 includes: Step 6.1, calculate the significance difference value d(F(i, j), F(i1, j1)) between different points at coordinates (i, j) and (i1, j1) in the initial hyperspectral spatial-spectral joint feature map F ∈ R M×N as follows: where c represents the weight factor, and d position is the Euclidean distance between the coordinates (i, j) and (i1, j1); Step 6.2, according to the following formula, accumulate the sum of the saliency difference values between a point and the global image as the saliency value result S(i, j) at the point (i, j), and sequentially traverse all the pixels in the image to obtain the initial saliency value image S; Step 6.3, set the threshold th according to the following formula, mark the elements in the initial saliency value image S that are greater than the threshold as significantly abnormal elements, and update the saliency value with the Euclidean distance between each point and the nearest adjacent significantly abnormal element as the weight factor to obtain a new saliency image In the formula, is the new significant value after background suppression based on the context before and after, and d near is the Euclidean distance between the point (i, j) and the nearest significant abnormal element in the vicinity; Step 6.4, based on the new significant image Judge each element type and output the detection result.

Images