A classification method for warships and cargo ships based on multi-feature screening

Abstract

The present invention relates to a classification method for ships and cargo ships based on multi-feature screening. The method is mainly applicable to the classification of ships and cargo ships of conventional broadband coherent warning radar under a dual-platform collaborative system. The main process is: firstly, the strong scattering center structure feature of a one-dimensional image at each viewing angle is extracted; the initial matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle is constructed; then, a combined feature matrix is ​​constructed by standardization processing and matrix merging; the optimal projection direction of the combined feature matrix is ​​calculated based on an improved two-dimensional Fisher linear discriminant analysis algorithm (A2DFLD); finally, a distance discriminant function is constructed to classify and identify the target type. The method provided by the present invention has the characteristics of simple engineering implementation and sufficient theoretical basis of the method used, and the classification and identification accuracy of ships and cargo ships under the dual-platform viewing angle is improved by more than 5% compared with the classification and identification accuracy of ships and cargo ships under a single viewing angle.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and is mainly applicable to the classification and recognition of warships and cargo ships by conventional broadband coherent warning radars in a collaborative system. Background Art

[0002] Research has shown that one-dimensional images can reflect the fine geometric structure characteristics of radar targets on the radar line of sight. Analysis of the structures of ships and cargo ships found that there are significant differences in physical structural characteristics between ships and cargo ships. The differences in physical structural characteristics lead to obvious differences in the one-dimensional image echoes of ships and cargo ships, such as the following: Figure 2 , Attachment Figure 3 , Attachment Figure 4 , Attachment Figure 5 Based on the one-dimensional images of each viewing angle, the classification and identification of ships and cargo ships can be realized through feature fusion, and the classification and identification information of ships and cargo ships can be used to achieve rough classification of radar targets and improve the recognition accuracy. In conjunction with the transformation of existing detection equipment, the collaborative target recognition capability can be greatly improved on the basis of existing detection equipment.

[0003] At present, many one-dimensional image technology research is based on a single-view one-dimensional image, such as the maximum correlation coefficient voting criterion sub-band fusion recognition algorithm proposed in the Journal of Xi'an University of Posts and Telecommunications in May 2013, "A New Target Recognition Algorithm Based on High-resolution Range Image Sub-band Fusion", which realizes radar target recognition under a single view. Single-view one-dimensional image target recognition is greatly affected by the target posture, which limits the engineering application of one-dimensional image target recognition.

[0004] Different from the methods proposed in other documents, the present invention aims at realizing the coarse classification of radar targets by a classification and identification method of ships and cargo ships through multi-feature screening and feature fusion based on a one-dimensional image from multiple perspectives. Summary of the invention

[0005] The purpose of the present invention is to provide a method for classifying warships and cargo ships based on multi-feature screening, so as to effectively realize the coarse classification of radar targets in a collaborative system.

[0006] The technical solution to realize the present invention is:

[0007] First, extract the strong scattering center structure features of the one-dimensional image at each viewing angle; construct the initial feature matrix of the strong scattering center structure feature base of the one-dimensional image at each viewing angle; then construct the combined feature matrix through flat order normalization and matrix merging; calculate the optimal projection direction of the combined feature matrix based on the improved two-dimensional Fisher linear discriminant analysis algorithm (A2DFFLD); finally, construct the distance discriminant function to classify and identify the target type. The specific steps include:

[0008] Step 1: Extract the strong scattering center structure features of the one-dimensional image at each viewing angle;

[0009] Step 2: Construct the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle;

[0010] Step 3: Construct the combined feature matrix;

[0011] Step 3-1: Performing flat-order normalization processing on the initial characteristic matrix of the strong scattering center structure characteristic basis at each viewing angle;

[0012] Step 3-2: Merge the initial matrix of the strong scattering center structure characteristic basis after the flat sequence normalization process to construct a combined characteristic matrix;

[0013] Step 4: Calculate the optimal projection direction of the combined feature matrix based on the improved two-dimensional Fisher linear discriminant analysis algorithm (A2DFLD);

[0014] Step 5: Construct a distance discriminant function to classify and identify target types.

[0015] The present invention performs flat order normalization processing on the characteristic basis initial matrix of the strong scattering center structure under each viewing angle, which can effectively eliminate the influence caused by the data asymmetry generated by combining two basis initial characteristic matrices, and can accurately and effectively eliminate the influence of the one-dimensional image posture sensitivity caused by different viewing angles. Based on the improved two-dimensional Fisher linear discriminant analysis algorithm (A2DFFLD) to calculate the best projection direction method of the combined characteristic matrix, the inter-class discrete matrix and the intra-class discrete matrix model are constructed, which makes it easier to represent the differences between samples of different categories and effectively classify samples of the same category. Through the present invention, it is possible to realize the effective classification of ships and cargo ships in various motion states within the visual range in the combat system of the maritime ship formation system, and the classification and recognition accuracy of ships and cargo ships under multi-view one-dimensional images is improved by 5% based on the classification and recognition accuracy of ships and cargo ships under single-view one-dimensional images.

[0016] The proposal and engineering realization of the present invention have high popularization and application value in the field of radar target recognition.

[0017] The present invention is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a data processing flow chart of the present invention.

[0019] Figure 2 It is a schematic diagram of the physical structure of the cargo ship of the present invention.

[0020] Figure 3 It is a schematic diagram of a one-dimensional image of a cargo ship according to the present invention.

[0021] Figure 4 It is the physical structure and schematic diagram of the ship of the present invention.

[0022] Figure 5 It is a schematic diagram of a one-dimensional image of a ship according to the present invention.

[0023] Figure 6 It is a flow chart of the one-dimensional image waveform structure feature extraction and processing of the target area under various viewing angles of the present invention.

[0024] Figure 7 It is a schematic diagram of the waveform structure characteristics of the one-dimensional image of the target area at various viewing angles of the present invention. DETAILED DESCRIPTION

[0025] The present invention provides a method for classifying warships and cargo ships based on multi-feature screening, and the preferred implementation steps are shown in the attached Figure 1 :

[0026] Step 1: Extract the strong scattering center structure features of the one-dimensional image at each viewing angle as follows:

[0027] (1) Preprocessing of the target one-dimensional image data at each viewing angle, specifically:

[0028] 1) Calculate the kurtosis matrix K of the one-dimensional image, find the maximum element max(K) of the kurtosis matrix and its corresponding one-dimensional image, and count the remaining one-dimensional image set {x i} and the outlier kurtosis set {k i}:

[0029]

[0030] Where N is the number of one-dimensional images, X i is the remaining i-th one-dimensional image sample data, μ i is the sample mean, σ i is the sample standard deviation, E(X i -μ i ) 4 is the fourth-order center distance of the i-th one-dimensional image sample data. If k i If it is not positive, the i-th one-dimensional image is considered to be an abnormal one-dimensional image and the sample data is set to 0.

[0031] 2) The one-dimensional image corresponding to the largest element of the kurtosis matrix is ​​used as the base image, and the minimum entropy spectrum estimation criterion is used for alignment, and the aligned one-dimensional image is incoherently accumulated.

[0032] I(x k )=-logp k

[0033]

[0034] where x k is a one-dimensional image sample data, I(x k ) is the amount of information, X is a discrete random variable, P k It is event X=x k The probability of occurrence, It is a one-dimensional image sample data set.

[0035]

[0036] where s i is the incoherent accumulation result of the ith distance unit after alignment, and N is the number of one-dimensional images.

[0037] (2) Extraction of target one-dimensional image region under each viewing angle, specifically:

[0038] Calculate the mean of the first and last eighth parts of the incoherently accumulated one-dimensional image of the target at each viewing angle, take 1.5 times of the minimum mean as the target one-dimensional image area extraction threshold, and the part exceeding the threshold is the target one-dimensional image area at each viewing angle.

[0039] (3) Extraction of waveform structural features of the target one-dimensional image region at each viewing angle, as shown in the attached figure. Figure 6 , specifically:

[0040] 1) Normalize the target one-dimensional image area at each viewing angle:

[0041]

[0042] Where P(m, n) is the sample data of the target one-dimensional image area, max(P(m, n)) is the global maximum of the sample data, min(P(m, n)) is the global minimum of the sample data, and P′(m, n) is the sample data after normalization and initialization.

[0043] 2) The normalized target one-dimensional image area is divided into 10 equal parts according to the distance unit {Q i}, i=1, 2,…, 10.

[0044] 3) Set the strong scattering point extraction threshold according to the method in (2) and count the number of strong scattering points M.

[0045] 4) Count the position and amplitude information of M strong scattering points.

[0046] 5) Carry out statistics on the number of strong scattering points, the position distribution statistics of strong scattering points and the amplitude distribution statistics of strong scattering points in equally divided areas respectively.

[0047] 6) Based on 3), 4), and 5), the nearest neighbor algorithm is used to extract waveform structural features including waveform effective width, waveform symmetry, waveform concentration, waveform kurtosis, number of waveform peaks, peak leading edge steepness, peak trailing edge steepness, and peak shape statistics, as shown in the attached figure. Figure 7 .

[0048] 7) Based on the extraction of the one-dimensional image intensity scattering center structure features of the target at each viewing angle, an initial matrix F of the one-dimensional image intensity scattering center structure features at each viewing angle is constructed i ∈R 1×8

[0049] F i =[y1,y2,y3,y4,y5,y6,y7,y8]

[0050] Among them, y1 represents the effective width of the waveform, y2 represents the symmetry of the waveform, y3 represents the concentration of the waveform, y4 represents the kurtosis of the waveform, y5 represents the number of waveform peaks, y6 represents the steepness of the leading edge of the kurtosis, y7 represents the steepness of the trailing edge of the kurtosis, and y8 represents the peak shape.

[0051] Step 2: construct the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle, the method is as follows:

[0052] In the initial matrix of the strong scattering center structure feature of the one-dimensional image at each viewing angle, two initial matrices with similar postures and high signal-to-noise ratio are selected to construct the initial feature matrix F of the strong scattering center structure feature base of the one-dimensional image at each viewing angle i ∈R 8×9 , as follows:

[0053]

[0054] where x i represents the position of the i-th feature at each viewing angle, y i represents the value of the i-th feature at each viewing angle, d ij Represents the Mahalanobis distance between the i-th feature and the j-th feature at each viewing angle.

[0055] Step 3: Construct the combined feature matrix as follows:

[0056] Step 3-1 performs flat order normalization processing on the initial feature matrix of the strong scattering center structure feature basis under each viewing angle, specifically:

[0057] (1) Calculate the mean matrix of the initial characteristic matrix of the strong scattering center structure characteristic basis of the one-dimensional image at each viewing angle

[0058]

[0059] Among them, F iIt is the initial characteristic matrix of the strong scattering center structure characteristic basis of the one-dimensional image at each viewing angle.

[0060] (2) Calculate the mean of the variance matrix of the initial characteristic matrix of the strong scattering center structure characteristic basis of the one-dimensional image at each viewing angle

[0061]

[0062] Among them, F i is the initial characteristic matrix of the strong scattering center structure characteristic basis of the one-dimensional image at each viewing angle, and μ is the mean matrix of the initial characteristic matrix of the strong scattering center structure characteristic basis of the one-dimensional image at each viewing angle.

[0063] (3) The initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle is subjected to flat order normalization processing, specifically:

[0064]

[0065] Among them, F i is the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle, μ is the mean matrix of the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle, σ is the mean of the variance matrix of the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle, and θ is the attitude angle at each viewing angle.

[0066] Step 3-2 merges the initial matrix of the strong scattering center structure characteristic basis after the flat order normalization process to construct a combined characteristic matrix, specifically:

[0067] The initial characteristic matrices of the two strong scattering center structure characteristic bases under the dual platform perspective are merged in a front-to-back parallel manner to construct a combined characteristic matrix

[0068] A=[F1′F2′]

[0069] Step 4: Calculate the optimal projection direction of the combined feature matrix. The optimal projection direction of the combined feature matrix is ​​calculated based on the improved two-dimensional Fisher linear discriminant analysis algorithm (A2DFLD), specifically:

[0070] (1) According to the two-dimensional Fisher linear discriminant analysis algorithm, the average combination feature matrix and the overall average combination feature matrix are calculated

[0071]

[0072]

[0073]

[0074] Where C is the total number of sample categories, A irepresents the combined feature matrix of the i-th type of target, represents the average combined feature matrix of the i-th type of target, represents the overall average combined feature matrix, ω i Represents the target identity.

[0075] (2) Calculate the difference distance of the combined feature matrix, specifically:

[0076] In a given data sample set X = {A1, A2, A3, ..., A N}, A i The category label is denoted as L i , and i = {1,2,…,C}, where C is the total number of sample categories, the constructed variant difference distance is as follows:

[0077]

[0078] Where d(A i ,A j )=||A i -A j ‖ represents the Euclidean distance between two data points, and the parameter β is the mean of the Euclidean distance between sample points, which is used to control D(A i ,A j )’s growth rate, and the parameter α is a constant.

[0079] (3) According to the two-dimensional Fisher linear discriminant analysis algorithm, an improved inter-class discrete matrix S is constructed b and the intra-class discrete matrix S w

[0080]

[0081]

[0082] Where C is the total number of sample categories, A i Represents the combined feature matrix of the i-th type of target, A j represents the combined feature matrix of the j-th target, represents the average combined feature matrix of the i-th type of target, represents the overall average combined feature matrix, f(A i ,A j )=1 / D(A i ,A j ) is the inter-class penalty factor, g(A i ,A j )=D(A i ,A j ) is the intra-class penalty factor.

[0083] (4) Calculate the optimal projection direction of the combined feature matrix based on the two-dimensional Fisher linear discriminant analysis algorithm

[0084]

[0085] Where ω is the optimal projection direction of the combined feature matrix, tr() represents the trace of the matrix, S b and S w They are the inter-class discrete matrix and the intra-class discrete matrix respectively.

[0086] According to linear algebra theory, the optimal projection direction ω opt That is the matrix S w -1 S b The maximum eigenvalue λ max The corresponding eigenvector is

[0087] S w -1 S b ω opt =λ max ω opt

[0088] It can be obtained that the optimal projection of the combined feature matrix under dual-platform collaborative detection is the eigenvector ω corresponding to the first two largest eigenvalues opt =[ω1,ω2].

[0089] Step 5: Construct a distance discriminant function to classify and identify target types. The method is as follows:

[0090] L=ω1 2 -ω2 2

[0091] When L is greater than zero, the target type is a ship; when L is less than zero, the target is a cargo ship; when L is equal to zero, classification and identification cannot be performed.

Claims

1. A method for classifying warships and cargo ships based on multi-feature screening, characterized in that: Step 1: Extract the strong scattering center structure features of the one-dimensional image at each viewing angle; Step 2: Construct the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle. First, calculate the Mahalanobis distance of each feature in the initial matrix [y1, y2, y3, y4, y5, y6, y7, y8] of the strong scattering center structure feature of the one-dimensional image at each viewing angle, where y1 represents the effective width of the waveform, y2 represents the symmetry of the waveform, y3 represents the concentration of the waveform, y4 represents the kurtosis of the waveform, y5 represents the number of waveform peaks, y6 represents the steepness of the leading edge of the kurtosis, y7 represents the steepness of the trailing edge of the kurtosis, and y8 represents the peak shape. Construct the initial feature matrix of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle where x m represents the position of the mth feature at each viewing angle, y m represents the value of the mth feature at each viewing angle, d mn Represents the Mahalanobis distance between the mth feature and the nth feature at each viewing angle; Step 3: Construct the combined feature matrix; Step 3-1: Perform a flat order normalization process on the initial feature matrix of the strong scattering center structure feature basis at each viewing angle. First, calculate the initial feature matrix F of the strong scattering center structure feature basis of the one-dimensional image at each viewing angle. i Then calculate the initial characteristic matrix F of the strong scattering center structure characteristic basis of the one-dimensional image at each viewing angle i The mean σ of the variance matrix is ​​calculated Where θ is the one-dimensional image attitude angle at each viewing angle; Step 3-2: Merge the initial characteristic matrix of the strong scattering center structure characteristic basis after the flat sequence normalization process to construct a combined characteristic matrix; Step 4: Calculate the optimal projection direction of the combined feature matrix based on the improved two-dimensional Fisher linear discriminant analysis algorithm; Step 5: Construct a distance discriminant function to classify and identify target types.

2. The method for classifying warships and cargo ships based on multi-feature screening according to claim 1 is characterized by: The step 4 also includes: Where ω is the optimal projection direction of the combined feature matrix, tr() represents the trace of the matrix, S b and S w are the inter-class discrete matrix and the intra-class discrete matrix respectively; based on the two-dimensional Fisher linear discriminant analysis algorithm, S b and S w To make improvements: Where C is the total number of sample categories, A i Represents the combined feature matrix of the i-th type of target, A k represents the combined feature matrix of the k-th target, represents the average combined feature matrix of the i-th type of target, represents the overall average combined feature matrix, f(A i ,A k ) is the inter-class penalty factor, g(A i ,A k ) is the intra-class penalty factor.

Images