Masked target detection and effect evaluation method adopting internal and external window spectrum unmixing mechanism

Abstract

The invention discloses a masked target detection and effect evaluation method adopting an internal and external window spectrum unmixing mechanism, which comprises the following steps: step 1, establishing an internal and external window detection mechanism; step 2, respectively obtaining end member pure spectrum dictionaries of an internal window background region and an external window external field region through a VCA unsupervised algorithm; step 4, obtaining an end member spectrum dictionary expression of each pixel; and step 5, constructing a detector to carry out abnormal target detection. The step 3 comprises the steps of: obtaining an abundance estimation matrix of each pixel based on the hyperspectral data and the pure end member spectrum matrix. And 6, setting a detection threshold value to obtain a detection and analysis result. According to the method disclosed by the invention, an inner and outer window detection mechanism is designed, spectral characteristics of an inner window background area and an outer window environment area are fully mined, dictionary expression is carried out on each pixel through a designed target detector, and the purpose of accurate abnormal target detection is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of hyperspectral earth remote sensing observation, obscured target survey and evaluation, and target characteristic analysis, and particularly relates to an obscured target detection and effect evaluation method in this field using an inner and outer window spectral unmixing mechanism. Background Art

[0002] Hyperspectral remote sensing organically combines spectral technology and imaging technology. It can not only detect the two-dimensional geometric spatial information of the target, but also the one-dimensional spectral information of the target. It enables the detection of objects that were originally undetectable through multi-spectral remote sensing.

[0003] Hyperspectral remote sensing has better spectral resolution than multispectral remote sensing, and can obtain continuous band images in the visible to near-infrared region. Each pixel in the image can form a smooth spectral curve with characteristic information. Through hyperspectral images, not only can we obtain images with target space and geometric relationship information, but also the spectral characteristics of the target. This feature is very suitable for detecting ground-covered targets and evaluating their camouflage effects.

[0004] The masked target detection technology using hyperspectral remote sensing images has gradually become an important means of battlefield reconnaissance and exploration, and the target detection technology based on the spectral dimension has also received more and more attention. At present, hyperspectral imaging technology has been successfully applied to the fields of remote sensing and aerospace military reconnaissance. Hyperspectral imagers can simultaneously image the same target in continuous spectral segments, which can directly reflect the spectral characteristics of the observed object and even the composition of the surface material of the object. Studies have shown that hyperspectral data can obtain an accurate relationship between space detection information and actual ground targets. Through the detected spectral characteristic curve, the material end member spectrum and its abundance estimation of each pixel can be calculated, so as to distinguish the difference between the natural background and the military target, and judge the nature and type of the target. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art in which the algorithm for abnormal target detection is based on the difference in target and background distribution, such as high false alarm rate, and inaccurate data of manually labeled pure spectrum training sets in supervised deep learning methods, and to provide a method for masked target detection and effect evaluation using an inner and outer window spectral unmixing mechanism.

[0006] The present invention adopts the following technical solution:

[0007] A method for detecting and evaluating a masked target using an inner and outer window spectral unmixing mechanism, the improvement of which is that it comprises the following steps:

[0008] Step 1: Establish internal and external window detection mechanism:

[0009] Step 11, initialize the sizes of the inner and outer windows, delete possible abnormal target pixels by masking the inverted inner window size area, and obtain the background area data cube of the outer window size;

[0010] Step 12, obtaining the environmental area data cube of the size of the external window;

[0011] Step 13, combining the edge area padding 0 method in deep learning, completing the data filling of the edge area, setting the step size, sliding the detection frame of the outer window size until the entire image is searched, and establishing an unsupervised inner and outer window detection mechanism;

[0012] Step 2: Use the VCA unsupervised algorithm to obtain the endmember pure spectrum dictionary of the inner window background area and the outer window field area respectively:

[0013] Step 21, using the endmember extraction algorithm, setting the number of endmember extractions, and extracting the pure spectrum dictionary E of the background area after the mask is inverted 背景 ;

[0014] Step 22, using the endmember extraction algorithm, setting the number of endmember extractions, and extracting the pure spectrum dictionary E of the environmental area 环境 ;

[0015] Step 3, using the non-negative least squares method NNLS, the abundance estimation matrix of each pixel is obtained based on the hyperspectral data and the pure end member spectrum matrix:

[0016] Step 31, using a spectral unmixing algorithm, based on the masked background area and the corresponding pure spectrum matrix, calculate the abundance estimation matrix A of each pixel in the sliding window area for the background pure spectrum matrix 背景 ;

[0017] Step 32, using a spectral unmixing algorithm, according to the environmental area and the corresponding pure spectrum matrix, calculate the abundance estimation matrix A of each pixel in the sliding window area for the environmental pure spectrum matrix 环境 ;

[0018] Step 4, obtain the endmember spectrum dictionary expression of each pixel:

[0019] Step 41, estimating the matrix A based on the acquired background area abundance 背景 and the corresponding pure spectrum dictionary E 背景 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 背景重组 ;

[0020] Step 42, estimating the matrix A based on the acquired environmental area abundance 环境 and the corresponding pure spectrum dictionary E 环境, calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 环境重组 ;

[0021] Step 5: Construct a detector to detect abnormal targets:

[0022] Step 51, multiplying the endmember spectrum dictionary expression corresponding to each pixel in the background and environment area with the abundance estimation matrix to obtain the endmember vector of the pixel;

[0023] D 背景重组i,j =E 背景 *A 背景i,j

[0024] H 环境重组i,j =E 环境 *A 环境i,j

[0025] In the above formula, i represents the i-th row, j represents the j-th column;

[0026] Step 52, subtract the reflectance of the corresponding pixel in the original pre-processed hyperspectral data from the end-member vectors of the background and environment regions to obtain the expression noise difference:

[0027]

[0028] In the above formula, x i,j represents the characteristic expression of the element in the i-th row and j-th column of the original data; β represents the adjustment factor;

[0029] Step 6: Set the detection threshold and obtain the detection and analysis results:

[0030] Step 61, performing unsupervised abnormal target detection on the entire image, obtaining a detection result with a value range of 0 to 1, setting a segmentation threshold, and obtaining a segmented binary image of the target and the background;

[0031] Step 62, obtaining the corresponding relationship between the camouflage level and the ground object segmentation threshold through different ground object segmentation thresholds, and then obtaining the camouflage level analysis result of the abnormal concealed target.

[0032] Furthermore, in step 11, the sizes of the inner and outer windows are initialized to be odd numbers.

[0033] Further, in steps 21 and 22, the endmember extraction algorithm includes a VCA vertex component analysis algorithm.

[0034] The beneficial effects of the present invention are:

[0035] The method disclosed in the present invention, by designing an inner and outer window detection mechanism, fully exploits the spectral characteristics of the inner window background area and the outer window environment area, and performs dictionary expression on each pixel through the designed target detector, so as to achieve the purpose of accurate abnormal target detection, and solves the shortcomings of high false alarm rate in traditional algorithms for abnormal target detection through the difference in target and background distribution, and the shortcomings of inaccurate manually labeled pure spectrum training set data and high labor cost in supervised deep learning methods. Compared with traditional methods, the results are more accurate and scientific. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0037] Figure 2 It is a detailed processing flow diagram of the method of the present invention. DETAILED DESCRIPTION

[0038] 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 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 intended to limit the present invention.

[0039] Example 1: This example discloses a method for detecting and evaluating obscured targets using an inner and outer window spectral unmixing mechanism. This method can draw on the idea of ​​the RX detector to establish an inner and outer window spectral unmixing mechanism, and achieve abnormal target detection by recursively accessing the entire image. By adjusting parameters, the detection results are more accurate and scientific. Figure 1 As shown, the following steps are included:

[0040] Step 1: Establish internal and external window detection mechanism:

[0041] Step 11, initialize the sizes of the inner and outer windows, use the inner window as the target detection buffer, and delete possible abnormal target pixels by masking the inner window size area to obtain the background area data cube of the outer window size;

[0042] Step 12, the outer window is taken as a mixed area containing the target and the background, and an environment area data cube of the outer window size is obtained;

[0043] Step 13, combining the edge area padding 0 method in deep learning, completing the data filling of the edge area, setting the step size, sliding the detection frame of the outer window size until the entire image is searched, and establishing an unsupervised inner and outer window detection mechanism;

[0044] Step 2: Use the VCA unsupervised algorithm to obtain the endmember pure spectrum dictionary of the inner window background area and the outer window field area respectively:

[0045] Step 21, using an endmember extraction algorithm, including but not limited to a VCA vertex component analysis algorithm, setting the number of endmember extractions, and extracting a pure spectrum dictionary E of the background area after the mask is inverted 背景 ;

[0046] Step 22, using an end member extraction algorithm, including but not limited to a VCA vertex component analysis algorithm, setting the number of end member extractions, and extracting a pure spectrum dictionary E of the environmental region 环境 ;

[0047] Step 3, using the non-negative least squares method NNLS, the abundance estimation matrix of each pixel is obtained based on the hyperspectral data and the pure end member spectrum matrix:

[0048] Step 31, using a spectral unmixing algorithm, including but not limited to the NNLS non-negative least squares method, based on the masked background area and the corresponding pure spectrum matrix, calculate the abundance estimation matrix A of each pixel in the sliding window area for the background pure spectrum matrix 背景 ;

[0049] Step 32, using a spectral unmixing algorithm, including but not limited to the NNLS non-negative least squares method, according to the environmental area and the corresponding pure spectrum matrix, calculate the abundance estimation matrix A of each pixel in the sliding window area for the environmental pure spectrum matrix 环境 ;

[0050] Step 4, obtain the endmember spectrum dictionary expression of each pixel:

[0051] Step 41, estimating the matrix A based on the acquired background area abundance 背景 and the corresponding pure spectrum dictionary E 背景 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 背景重组 ;

[0052] Step 42, estimating the matrix A based on the acquired environmental area abundance 环境 and the corresponding pure spectrum dictionary E 环境 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 环境重组 ;

[0053] Step 5: Construct a detector to detect abnormal targets:

[0054] Step 51, multiplying the endmember spectrum dictionary expression corresponding to each pixel in the background and environment area with the abundance estimation matrix to obtain the endmember vector of the pixel;

[0055] D 背景重组i,j =E 背景 *A 背景i,j

[0056] H 环境重组i,j =E 环境 *A 环境i,j

[0057] In the above formula, i represents the i-th row, j represents the j-th column;

[0058] Step 52, subtract the reflectance of the corresponding pixel in the original pre-processed hyperspectral data from the end-member vectors of the background and environment regions to obtain the expression noise difference:

[0059]

[0060] In the above formula, x i,j It represents the characteristic expression of the element in the i-th row and j-th column of the original data; β represents the adjustment factor, which is used to control the calculation result within the recognizable range;

[0061] Step 6: Set the detection threshold and obtain the detection and analysis results:

[0062] Step 61, performing unsupervised abnormal target detection on the entire image, obtaining a detection result with a value range of 0 to 1, setting a segmentation threshold, and obtaining a segmented binary image of the target and the background;

[0063] Step 62, obtaining the corresponding relationship between the camouflage level and the ground object segmentation threshold through different ground object segmentation thresholds, and then obtaining the camouflage level analysis result of the abnormal concealed target.

[0064] For ease of understanding, the method of the present invention is further described in detail below. Figure 2 As shown, the specific steps include:

[0065] Step 201: Establishing an internal and external window detection mechanism, the specific implementation process is as follows:

[0066] (11): Drawing on the RX anomaly detection idea, initialize the inner and outer window sizes and set them to odd numbers. The inner window is used as the target detection buffer. By "masking the inverse" of the inner window size area, it is multiplied with the original outer window area data matrix to obtain the background area matrix with all inner window data 0, and obtain the background area data cube of the outer window size;

[0067] (12): Initialize the size of the inner and outer windows, set them to an odd number, use the outer window as a mixed area containing the target and the background, and obtain the environmental area data cube of the outer window size;

[0068] (13): Combined with the edge area padding 0 method in deep learning, the edge area data is filled, the step size is set, and the detection box of the outer window size is slid until the entire image is searched, and an unsupervised inner and outer window detection mechanism is established;

[0069] Step 202: Obtain the pure spectrum matrix corresponding to the inner and outer windows. The specific implementation process is:

[0070] (21): Through the end member extraction algorithm, including but not limited to the VCA vertex component analysis method, the number of end member extractions is set, and the pure spectrum matrix E of the background area after "mask inversion" is extracted 背景 ;

[0071] (22): Through the end member extraction algorithm, including but not limited to the VCA vertex component analysis method, the number of end member extractions is set to extract the pure spectral matrix E of the environmental area 环境 .

[0072] Step 203: Calculate the abundance estimate of the pure spectrum corresponding to the background area. The specific implementation process is:

[0073] (31): Through a spectral unmixing algorithm, including but not limited to the NNLS non-negative least squares method, based on the background area data and the corresponding pure spectrum matrix obtained by (21), the abundance estimation matrix A of each pixel in the sliding window area for the background pure spectrum matrix is ​​calculated. 背景 ;

[0074] Step 204: Calculate the abundance estimate of the pure spectrum corresponding to the environmental area. The specific implementation process is:

[0075] (41): Through a spectral unmixing algorithm, including but not limited to the NNLS non-negative least squares method, according to the environmental area and the corresponding pure spectral matrix obtained by (22), the abundance estimation matrix A of each pixel in the sliding window area for the environmental pure spectral matrix is ​​calculated. 环境 .

[0076] Step 205: Calculate the dictionary expression of each pixel in the background area. The specific implementation process is:

[0077] (51): Estimate the matrix A based on the obtained background area abundance 背景 and the corresponding pure spectrum dictionary E 背景 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 背景重组 ;

[0078] (52): Multiply the endmember dictionary expression and abundance estimation matrix corresponding to each pixel in the background area, and obtain the endmember vector of the pixel according to the following formula.

[0079] D 背景重组i,j =E 背景 *A 背景i,j

[0080] Step 206: Calculate the dictionary expression of each pixel in the environment area. The specific implementation process is:

[0081] (61): Estimate the matrix A based on the obtained environmental area abundance 环境 and the corresponding pure spectrum dictionary E 环境 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 环境重组 ;

[0082] (62): Multiply the endmember dictionary expression corresponding to each pixel in the environmental area and the abundance estimation matrix, and obtain the endmember vector of the pixel according to the following formula.

[0083] H 环境重组i,j =E 环境 *A 环境i,j

[0084] Step 207: Construct an abnormal target detector. The specific implementation process is as follows:

[0085] (71): Subtract the reflectance of the corresponding pixel in the original preprocessed hyperspectral data from the end-member vectors of the background and ambient regions to obtain the expression noise difference;

[0086]

[0087] (72): Perform unsupervised abnormal target detection on the entire image, obtain a detection result with a value range of 0 to 1, set a segmentation threshold, and obtain a segmented binary image of the target and background;

[0088] Step 208: Obtain the detection and analysis results. The specific implementation process is as follows:

[0089] (81): By using different ground object segmentation thresholds, the corresponding relationship between the camouflage level and the ground object segmentation threshold is obtained, and the camouflage level analysis result of the abnormal target is obtained.

Claims

1. A method for detecting and evaluating obscured targets using an inner and outer window spectral unmixing mechanism, characterized in that: The steps include: Step 1: Establish internal and external window detection mechanism: Step 11, initialize the sizes of the inner and outer windows, delete possible abnormal target pixels by masking the inverted inner window size area, and obtain the background area data cube of the outer window size; Step 12, obtaining the environmental area data cube of the size of the external window; Step 13, combining the edge area padding 0 method in deep learning, completing the data filling of the edge area, setting the step size, sliding the detection frame of the outer window size until the entire image is searched, and establishing an unsupervised inner and outer window detection mechanism; Step 2: Use the VCA unsupervised algorithm to obtain the endmember pure spectrum dictionary of the inner window background area and the outer window field area respectively: Step 21, using the endmember extraction algorithm, setting the number of endmember extractions, and extracting the pure spectrum dictionary E of the background area after the mask is inverted 背景 ; Step 22, using the endmember extraction algorithm, setting the number of endmember extractions, and extracting the pure spectrum dictionary E of the environmental area 环境 ; Step 3, using the non-negative least squares method NNLS, the abundance estimation matrix of each pixel is obtained based on the hyperspectral data and the pure end member spectrum matrix: Step 31, using a spectral unmixing algorithm, based on the masked background area and the corresponding pure spectrum matrix, calculate the abundance estimation matrix A of each pixel in the sliding window area for the background pure spectrum matrix 背景 ; Step 32, using a spectral unmixing algorithm, according to the environmental area and the corresponding pure spectrum matrix, calculate the abundance estimation matrix A of each pixel in the sliding window area for the environmental pure spectrum matrix 环境 ; Step 4, obtain the endmember spectrum dictionary expression of each pixel: Step 41, estimating the matrix A based on the acquired background area abundance 背景 and the corresponding pure spectrum dictionary E 背景 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 背景重组 ; Step 42, estimating the matrix A based on the acquired environmental area abundance 环境 and the corresponding pure spectrum dictionary E 环境 , calculate the end member spectrum dictionary expression D of each pixel in the sliding window area 环境重组 ; Step 5: Construct a detector to detect abnormal targets: Step 51, multiplying the endmember spectrum dictionary expression corresponding to each pixel in the background and environment area with the abundance estimation matrix to obtain the endmember vector of the pixel; D 背景重组i,j =E 背景 *A 背景i,j H 环境重组i,j =E 环境 *A 环境i,j In the above formula, i represents the i-th row, j represents the j-th column; Step 52, subtract the reflectance of the corresponding pixel in the original pre-processed hyperspectral data from the end-member vectors of the background and environment regions to obtain the expression noise difference: In the above formula, x i,j Represents the characteristic expression of the element in the i-th row and j-th column of the original data; β represents the adjustment factor; Step 6: Set the detection threshold and obtain the detection and analysis results: Step 61, performing unsupervised abnormal target detection on the entire image, obtaining a detection result with a value range of 0 to 1, setting a segmentation threshold, and obtaining a segmented binary image of the target and the background; Step 62, obtaining the corresponding relationship between the camouflage level and the ground object segmentation threshold through different ground object segmentation thresholds, and then obtaining the camouflage level analysis result of the abnormal concealed target.

2. The method for detecting and evaluating obscured targets using the inner and outer window spectral unmixing mechanism according to claim 1, characterized in that: In step 11, the sizes of the inner and outer windows are initialized to odd numbers.

3. The method for detecting and evaluating obscured targets using the inner and outer window spectral unmixing mechanism according to claim 1, characterized in that: In steps 21 and 22, the endmember extraction algorithm includes a VCA vertex component analysis algorithm.

Images