Radar low-altitude target detection method based on multi-feature classification

Abstract

The present invention relates to a radar low-altitude target detection method based on multi-feature classification, comprising: extracting multiple features from a radar low-altitude RD image to obtain LBP texture features, grayscale features and grayscale maximum fluctuation features; forming a feature vector set F′, dividing the feature vector set F′ into a training data set and a test data set; training the training data set using a single-classification SVM algorithm to obtain a classification model; inputting the test data set into the classification model to perform target detection, and outputting a target detection result. The present invention mainly utilizes the difference between the target and the clutter background in the radar RD image to perform feature selection, constructs a feature training set that can distinguish between the target and the clutter, and then realizes accurate classification of low-altitude target pixels and clutter background pixels based on the single-classification SVM algorithm, and the effectiveness of the invention is proved by a large number of experiments.

Description

Technical Field

[0001] The invention relates to the field of radar application technology, and in particular to a radar low-altitude target detection method based on multi-feature classification. Background Art

[0002] In recent years, with the opening of low-altitude airspace, unmanned aerial vehicles have gradually been used in all walks of life, bringing great convenience to industry and daily life, but also posing a huge threat to low-altitude safety, such as leaks, public safety, aviation safety, etc. In order to meet the urgent needs of my country in the fields of civil aviation and military low-altitude penetration, how to achieve efficient and accurate low-altitude target detection has important research significance.

[0003] The low-altitude environment is complex. The scattering of low-altitude targets is weak. The flight altitude is relatively low, and they are easily hidden in clutter and difficult to extract, which affects the detection performance of the radar. For traditional CFAR detection methods, false alarms and missed detections are prone to occur. In recent years, deep learning has gradually been applied to radar low-altitude target detection, but the time complexity of deep learning is high, making it difficult to apply in practice. Summary of the invention

[0004] The purpose of the present invention is to provide a radar low-altitude target detection method based on multi-feature classification, which can efficiently detect low-altitude targets, especially low-altitude weak targets, and feature classification based on machine learning greatly improves the classification accuracy of target and clutter pixels, thereby improving target detection performance.

[0005] To achieve the above object, the present invention adopts the following technical solution: a radar low-altitude target detection method based on multi-feature classification, the method comprising the following steps in order:

[0006] (1) Obtaining the original radar low-altitude RD image to be detected;

[0007] (2) Perform multi-feature extraction on the original radar low-altitude RD image to be detected to obtain LBP texture features, grayscale features, and grayscale maximum fluctuation features;

[0008] (3) The radar RD image is subjected to sliding window processing to obtain an LBP texture feature vector dataset, a grayscale feature vector dataset, and a grayscale maximum fluctuation feature vector dataset. The LBP texture feature vector dataset, the grayscale feature vector dataset, and the grayscale maximum fluctuation feature vector dataset are normalized respectively to form a feature vector set F′. The feature vector set F′ is divided into a training dataset and a test dataset.

[0009] (4) Use the single-class SVM algorithm to train the training data set and obtain a classification model;

[0010] (5) Input the test data set into the classification model for target detection and output the target detection results.

[0011] The step (2) specifically comprises the following steps:

[0012] (2a) Calculate the LBP texture feature value of the local window area:

[0013] The local binary pattern LBP uses the gray value of the central pixel in the 3×3 window as the threshold, divides the pixels in the neighborhood, and records the pixels with a gray value greater than the central pixel as bright pixels, and the pixels with a gray value less than the central pixel as dark pixels. The texture structure information of the 3×3 local window area is represented as a matrix pattern:

[0014]

[0015] Among them, x c represents the gray value of the central pixel in the 3×3 local window area, and x(i,j) represents the gray value of any pixel in the 3×3 local window area; pixels with S value of 1 are recorded as bright pixels, and pixels with S value of 0 are recorded as dark pixels;

[0016] The texture intensity information of the 3×3 local window area is characterized by calculating the contrast between bright and dark pixels in the 3×3 local window area. The calculation formula of the LBP texture eigenvalue is:

[0017]

[0018] Among them, G p represents the gray value of the pth bright pixel in the 3×3 local window area, n1 represents the number of bright pixels in the 3×3 local window neighborhood; g q represents the gray value of the qth dark pixel in the 3×3 local window area, and n2 represents the number of dark pixels in the 3×3 local window neighborhood;

[0019] (2b) Calculate the grayscale eigenvalue of the local window area:

[0020] The grayscale feature statistics are performed on the local 3×3 window area, and the calculation formula of the grayscale feature value is:

[0021]

[0022] Among them, U k Indicates the kth maximum grayscale value in the local 3×3 window area, K is the number of maximum grayscale values, and 1K≤9;

[0023] (2c) Calculate the maximum grayscale fluctuation eigenvalue of the local window area:

[0024] The calculation formula of the maximum grayscale fluctuation eigenvalue is:

[0025]

[0026] Among them, I max Indicates the maximum grayscale value in the local 3×3 window area of ​​the radar RD image, I i ,i=1,2,...,9 represents the gray value of any pixel in the local 3×3 window area.

[0027] The step (3) specifically refers to:

[0028] The radar RD image is processed by sliding window to obtain three feature vector data sets, namely, the LBP texture feature vector data set F1 = {f1 m |m=1,2,...,M}, gray feature vector dataset F2={f2 m |m=1,2,...,M and the grayscale maximum fluctuation feature vector data set F3=f3mm=1,2,...,M, where M represents the total number of pixels contained in the radar image. The mth pixel is represented by a three-dimensional feature vector F m =[f1 m ,f2 m ,f3 m ],m=1,2,...,M to represent the three-dimensional feature vector F m The corresponding normalized eigenvector calculation formula is:

[0029]

[0030] Among them, F1 max 、F1 min Respectively represent the maximum and minimum values ​​in the LBP texture feature vector dataset F1, F2 max 、F2 min Respectively represent the maximum and minimum values ​​in the gray feature vector data set F2, F3 max 、F3 min They represent the maximum and minimum values ​​in the grayscale maximum fluctuation feature vector data set F3 respectively;

[0031] Then the three-dimensional feature vector set composed of M pixels in the radar RD image is expressed as:

[0032] F′={F′ m |m=1,2,...,M}

[0033] Among them, F′ is a three-dimensional feature vector set composed of the LBP texture feature vector data set F1, the grayscale feature vector data set F2, and the grayscale maximum fluctuation feature vector data set F3 after normalization processing, namely, the feature vector set F′;

[0034] A part of the feature vector set F′ that represents pure clutter pixels is selected, denoted as Z, and used as the training data set; the remaining pixel feature vector data set, denoted as ZT, is used as the test data set.

[0035] In step (4), the classification model specifically refers to:

[0036] For the training data set Z = {z1,z2,...,z N}, where z i ∈R 3 is a 3-dimensional real number space, the training data set Z contains N samples, and the hypersphere is constructed:

[0037]

[0038] s.t. ||F(z i )-c|| 2 ≤R 2 +ξ i , i=1,...,N,

[0039] ξ i ≥0, i=1, ..., N

[0040] Where R is the radius of the hypersphere, c is the center of the hypersphere, C is the penalty coefficient introduced to penalize samples falling outside the hypersphere, ξ i represents the relaxation factor, which is used to reduce the influence of singular points, F(z i ) is a nonlinear mapping that maps samples to a high-dimensional feature space;

[0041] The dual form of the simplified original optimization problem is:

[0042]

[0043]

[0044] 0≤α i ≤C, i=1,...,N

[0045] Among them, k(z i , z j ) is the Gaussian kernel function, and the optimal solution is calculated. The support vector sample set is SV, and its sample number is n sv , according to the false alarm rate optimization, the support vector z with the smallest modulus value in the support vector sample set SV is determined s Calculate the radius of the hypersphere as:

[0046]

[0047] Among them, F(z s ) represents the support vector z with the smallest modulus s is a nonlinear mapping, c is the center of the hypersphere, and N is the number of training samples.

[0048] The step (5) specifically means that according to the classification model, the detection and discrimination of any test sample zt∈ZT only requires calculating the distance from the sample feature vector to the center of the sphere, and the calculation formula is:

[0049]

[0050] In the formula, represents the square of the distance from any test sample zt to the center c of the hypersphere, F(zt) represents the nonlinear mapping of the test sample zt to the high-dimensional feature space, and N represents the number of training samples;

[0051] like If the target is larger than 0, it is judged as clutter background, otherwise it is judged as radar low-altitude target, and R is the radius of the hypersphere.

[0052] It can be seen from the above technical scheme that the beneficial effects of the present invention are as follows: the present invention can well solve the disadvantage that the target detection accuracy of radar low-altitude targets is reduced due to their weak scattering and easy hiding in the clutter background, and by performing multi-feature extraction on the original radar RD image, the constructed feature vector can more richly characterize the image pixel information, and the pixel feature vector is classified based on the single-classification SVM algorithm, which has a good classification effect and improves the detection rate of radar low-altitude targets in complex backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the method of the present invention;

[0054] Figure 2 is the selected original radar RD map to be detected;

[0055] Figure 3 This is the radar low-altitude target detection result diagram. DETAILED DESCRIPTION

[0056] like Figure 1 As shown, a radar low-altitude target detection method based on multi-feature classification includes the following steps in order:

[0057] (1) Obtain the original radar low-altitude RD image to be detected, such as Figure 2 As shown;

[0058] (2) Perform multi-feature extraction on the original radar low-altitude RD image to be detected to obtain LBP texture features, grayscale features, and grayscale maximum fluctuation features;

[0059] (3) The radar RD image is subjected to sliding window processing to obtain an LBP texture feature vector dataset, a grayscale feature vector dataset, and a grayscale maximum fluctuation feature vector dataset. The LBP texture feature vector dataset, the grayscale feature vector dataset, and the grayscale maximum fluctuation feature vector dataset are normalized respectively to form a feature vector set F′. The feature vector set F′ is divided into a training dataset and a test dataset.

[0060] (4) Use the single-class SVM algorithm to train the training data set and obtain a classification model;

[0061] (5) Input the test data set into the classification model for target detection and output the target detection results.

[0062] The step (2) specifically comprises the following steps:

[0063] (2a) Calculate the LBP texture feature value of the local window area:

[0064] Texture is a spatial distribution property that can reflect the grayscale of pixels in an image. The inherent properties of this spatial structure are mainly described by the correlation between neighboring pixels. Local Binary Pattern (LBP) is a description operator widely used to characterize the texture feature structure of local areas of images. It has the characteristics of simple statistical theory and low computational complexity. This operator mainly describes texture features by statistically analyzing the spatial distribution structure between pixels in the local window area. Generally, LBP mainly uses the grayscale value of the central pixel in a 3×3 window as the threshold, and divides the remaining pixels in the neighborhood. Pixels with a grayscale value greater than the central pixel are recorded as bright pixels, and those with a grayscale value less than the central pixel are recorded as dark pixels.

[0065] The local binary pattern LBP uses the gray value of the central pixel in the 3×3 window as the threshold, divides the pixels in the neighborhood, and records the pixels with a gray value greater than the central pixel as bright pixels, and the pixels with a gray value less than the central pixel as dark pixels. The texture structure information of the 3×3 local window area is represented as a matrix pattern:

[0066]

[0067] Among them, x c represents the gray value of the central pixel in the 3×3 local window area, and x(i,j) represents the gray value of any pixel in the 3×3 local window area; pixels with S value of 1 are recorded as bright pixels, and pixels with S value of 0 are recorded as dark pixels;

[0068] The texture intensity information of the 3×3 local window area is characterized by calculating the contrast between bright and dark pixels in the 3×3 local window area. The calculation formula of the LBP texture eigenvalue is:

[0069]

[0070] Among them, G p represents the gray value of the pth bright pixel in the 3×3 local window area, n1 represents the number of bright pixels in the 3×3 local window neighborhood; g q represents the gray value of the qth dark pixel in the 3×3 local window area, and n2 represents the number of dark pixels in the 3×3 local window neighborhood;

[0071] (2b) Calculate the grayscale eigenvalue of the local window area:

[0072] The grayscale characteristics of the radar RD image reflect the backscattering characteristics of the ground objects to the radar waves. Since the scattering characteristics of the target and the clutter background to the electromagnetic waves are different, the intensity of the scattered echo is also different. Generally, the grayscale characteristics of the clutter background are weaker, and the grayscale characteristics of the radar low-altitude targets are stronger.

[0073] The grayscale feature statistics are performed on the local 3×3 window area, and the calculation formula of the grayscale feature value is:

[0074]

[0075] Among them, I k Indicates the kth maximum grayscale value in the local 3×3 window area, K is the number of maximum grayscale values, and 1K≤9;

[0076] (2c) Calculate the maximum grayscale fluctuation eigenvalue of the local window area:

[0077] The present invention further introduces the grayscale maximum fluctuation parameter to quantitatively describe the texture feature contrast between the low-altitude radar target and the background. This feature reflects the maximum fluctuation variation of the texture in the local 3×3 window area. The calculation formula of the grayscale maximum fluctuation feature value is:

[0078]

[0079] Among them, I max Indicates the maximum gray value in the local 33 window area of ​​the radar RD image, I i ,i=1,2,...,9 represents the gray value at any pixel in the local 33 window area.

[0080] The step (3) specifically refers to:

[0081] The radar RD image is processed by sliding window to obtain three feature vector data sets, namely, the LBP texture feature vector data set F1 = {f1 m |m=1,2,...,M}, gray feature vector dataset F2={f2m |m=1,2,...,M and the grayscale maximum fluctuation feature vector data set F3f3mm=1,2,...,M, where M represents the total number of pixels contained in the radar image, then the mth pixel is represented by a three-dimensional feature vector F m =[f1 m ,f2 m ,f3 m ],m=1,2,...,M to represent the three-dimensional feature vector F m The corresponding normalized eigenvector calculation formula is:

[0082]

[0083] Among them, F1 max 、F1 min Respectively represent the maximum and minimum values ​​in the LBP texture feature vector dataset F1, F2 max 、F2 min Respectively represent the maximum and minimum values ​​in the gray feature vector data set F2, F3 max 、F3 min They represent the maximum and minimum values ​​in the grayscale maximum fluctuation feature vector data set F3 respectively;

[0084] Then the three-dimensional feature vector set composed of M pixels in the radar RD image is expressed as:

[0085] F′={F' m |m=1,2,...,M}

[0086] Among them, F′ is a three-dimensional feature vector set composed of the LBP texture feature vector data set F1, the grayscale feature vector data set F2, and the grayscale maximum fluctuation feature vector data set F3 after normalization processing, namely, the feature vector set F′;

[0087] A part of the feature vector set F′ that represents pure clutter pixels is selected, denoted as Z, and used as the training data set; the remaining pixel feature vector data set, denoted as ZT, is used as the test data set.

[0088] In step (4), the classification model specifically refers to:

[0089] For the training data set Z = {z1,z2,...,z N}, where z i ∈R 3 is a 3-dimensional real number space, the training data set Z contains N samples, and the hypersphere is constructed:

[0090]

[0091] st||F(zi )-c|| 2 ≤R 2 +ξ i , i=1,...,N,

[0092] ξ i ≥0, i=1, ..., N

[0093] Where R is the radius of the hypersphere, c is the center of the hypersphere, C is the penalty coefficient introduced to penalize samples falling outside the hypersphere, ξ i represents the relaxation factor, which is used to reduce the influence of singular points, F(z i ) is a nonlinear mapping that maps samples to a high-dimensional feature space;

[0094] The dual form of the simplified original optimization problem is:

[0095]

[0096]

[0097] 0≤α i ≤C, i=1,...,N

[0098] Among them, k(z i ,z j ) is the Gaussian kernel function, and the optimal solution is calculated. The support vector sample set is SV, and its sample number is n sv , according to the false alarm rate optimization, the support vector z with the smallest modulus value in the support vector sample set SV is determined s Calculate the radius of the hypersphere as:

[0099]

[0100] Among them, F(z s ) represents the support vector z with the smallest modulus s is a nonlinear mapping, c is the center of the hypersphere, and N is the number of training samples.

[0101] The step (5) specifically means that according to the classification model, the detection and discrimination of any test sample zt∈ZT only requires calculating the distance from the sample feature vector to the center of the sphere, and the calculation formula is:

[0102]

[0103] In the formula, represents the square of the distance from any test sample zt to the center c of the hypersphere, F(zt) represents the nonlinear mapping of the test sample zt to the high-dimensional feature space, and N represents the number of training samples;

[0104] like If the target is larger than 0, it is judged as clutter background, otherwise it is judged as radar low-altitude target, and R is the radius of the hypersphere.

[0105] Classifying the feature vector representation of all pixels in the radar RD image can determine the target pixels and background pixels, thereby completing the detection of the target, such as Figure 3 As shown, it can be seen from the detection results that the invention can accurately detect low-altitude targets by radar and improve the detection performance of the target, especially the detection of low-altitude weak targets.

[0106] In summary, the present invention mainly uses the difference between the target and the clutter background in the radar RD image to perform feature selection, and constructs a feature vector training set that can distinguish the target and the clutter, and then realizes the accurate classification of low-altitude target pixels and clutter background pixels based on the single-classification SVM algorithm, and the effectiveness of the invention is proved by a large number of experiments. The present invention can significantly improve the detection performance of the target in a complex clutter background, especially the low-altitude weak target.

Claims

1. A radar low-altitude target detection method based on multi-feature classification, characterized in that: The method comprises the following steps in order: (1) Obtaining the original radar low-altitude RD image to be detected; (2) Perform multi-feature extraction on the original radar low-altitude RD image to be detected to obtain LBP texture features, grayscale features, and grayscale maximum fluctuation features; (3) The radar RD image is subjected to sliding window processing to obtain an LBP texture feature vector dataset, a grayscale feature vector dataset, and a grayscale maximum fluctuation feature vector dataset. The LBP texture feature vector dataset, the grayscale feature vector dataset, and the grayscale maximum fluctuation feature vector dataset are normalized respectively to form a feature vector set F′. The feature vector set F′ is divided into a training dataset and a test dataset. (4) Use the single-class SVM algorithm to train the training data set and obtain a classification model; (5) Input the test data set into the classification model for target detection and output the target detection results; In step (4), the classification model specifically refers to: For the training data set Z = {z1,z2,...,z N }, where z i ∈R 3 is a 3-dimensional real number space, the training data set Z contains N samples, and the hypersphere is constructed: s.t.‖F(z i )-c‖ 2 ≤R 2 +ξ i ,i=1,...,N, x i ≥0,i=1,...,N Where R is the radius of the hypersphere, c is the center of the hypersphere, C is the penalty coefficient introduced to penalize samples falling outside the hypersphere, ξ i represents the relaxation factor, which is used to reduce the influence of singular points, F(z i ) is a nonlinear mapping that maps samples to a high-dimensional feature space; The dual form of the simplified original optimization problem is: 0≤α i ≤C,i=1,...,N Among them, k(z i ,z j ) is the Gaussian kernel function, and the optimal solution is calculated. The support vector sample set is SV, and its sample number is n sv , according to the false alarm rate optimization, the support vector z with the smallest modulus value in the support vector sample set SV is determined s Calculate the radius of the hypersphere as: Among them, F(z s ) represents the support vector z with the smallest modulus s is a nonlinear mapping, c is the center of the hypersphere, and N is the number of training samples.

2. The radar low-altitude target detection method based on multi-feature classification according to claim 1, characterized in that: The step (2) specifically comprises the following steps: (2a) Calculate the LBP texture feature value of the local window area: The local binary pattern LBP uses the gray value of the central pixel in the 3×3 window as the threshold, divides the pixels in the neighborhood, and records the pixels with a gray value greater than the central pixel as bright pixels, and the pixels with a gray value less than the central pixel as dark pixels. The texture structure information of the 3×3 local window area is represented as a matrix pattern: Among them, x c represents the gray value of the central pixel in the 3×3 local window area, and x(i, j) represents the gray value of any pixel in the 3×3 local window area; pixels with an S value of 1 are recorded as bright pixels, and pixels with an S value of 0 are recorded as dark pixels; The texture intensity information of the 3×3 local window area is characterized by calculating the contrast between bright and dark pixels in the 3×3 local window area. The calculation formula of the LBP texture eigenvalue is: Among them, G p represents the gray value of the pth bright pixel in the 3×3 local window area, n1 represents the number of bright pixels in the 3×3 local window neighborhood; g q represents the gray value of the qth dark pixel in the 3×3 local window area, and n2 represents the number of dark pixels in the 3×3 local window neighborhood; (2b) Calculate the grayscale eigenvalue of the local window area: The grayscale feature statistics are performed on the local 3×3 window area, and the calculation formula of the grayscale feature value is: Among them, I k Represents the kth maximum grayscale value in the local 3×3 window area, K is the number of maximum grayscale values, and 1≤K≤9; (2c) Calculate the maximum grayscale fluctuation eigenvalue of the local window area: The calculation formula of the maximum grayscale fluctuation eigenvalue is: Among them, I max Indicates the maximum grayscale value in the local 3×3 window area of ​​the radar RD image, I i , i=1,2,...,9 represents the grayscale value at any pixel in the local 3×3 window area.

3. The radar low-altitude target detection method based on multi-feature classification according to claim 1 is characterized in that: The step (3) specifically refers to: The radar RD image is processed by sliding window to obtain three feature vector data sets, namely, the LBP texture feature vector data set F1 = {f1 m |m=1,2,...,M}, gray feature vector dataset F2={f2 m |m=1,2,…,M and grayscale maximum fluctuation feature vector data set F3=f3mm=1,2,…,M, where M represents the total number of pixels contained in the radar image. Then the mth pixel is represented by a three-dimensional feature vector F m =[f1 m , f2 m , f3 m ], m = 1, 2, ..., M to represent the three-dimensional feature vector F m The corresponding normalized eigenvector calculation formula is: Among them, F1 max 、F1 min Respectively represent the maximum and minimum values ​​in the LBP texture feature vector dataset F1, F2 max 、F2 min Respectively represent the maximum and minimum values ​​in the gray feature vector data set F2, F3 max 、F3 min They represent the maximum and minimum values ​​in the grayscale maximum fluctuation feature vector data set F3 respectively; Then the three-dimensional feature vector set composed of M pixels in the radar RD image is expressed as: F′={F′m|m=1,2,…,M} Among them, F′ is a three-dimensional feature vector set composed of the LBP texture feature vector data set F1, the grayscale feature vector data set F2, and the grayscale maximum fluctuation feature vector data set F3 after normalization processing, namely, the feature vector set F′; A part of the feature vector set F′ that represents pure clutter pixels is selected, denoted as Z, and used as the training data set; the remaining pixel feature vector data set, denoted as ZT, is used as the test data set.

4. The radar low-altitude target detection method based on multi-feature classification according to claim 1 is characterized in that: The step (5) specifically means that according to the classification model, the detection and discrimination of any test sample zt∈ZT only requires calculating the distance from the sample feature vector to the center of the sphere, and the calculation formula is: In the formula, represents the square of the distance from any test sample zt to the center c of the hypersphere, F(zt) represents the nonlinear mapping of the test sample zt to the high-dimensional feature space, and N represents the number of training samples; like If the target is larger than 0, it is judged as clutter background, otherwise it is judged as radar low-altitude target, and R is the radius of the hypersphere.

Images