A radar one-dimensional range image multi-scale multi-kernel learning target recognition method

Abstract

The application discloses a radar one-dimensional range image multi-scale multi-kernel learning target recognition method, which is mainly suitable for sea surface ship target classification and identification of a shore-based sea police warning radar. The main process is as follows: in the training stage, firstly, multi-scale transformation is performed on one-dimensional range image training samples to form multi-scale transformation features under different scales; then, kernel space transformation processing is performed on the multi-scale transformation features to form kernel space transformations under different kernel functions; linear synthesis is performed on the kernel space transformations to construct a projection matrix; finally, the obtained training sample projection vectors are sent to an SVM classifier for training; in the test stage, firstly, the projection vectors of the training samples are obtained; then, the obtained training sample projection vectors are sent to the trained SVM classifier for testing. The nonlinear data is mapped to a high-dimensional linearly separable space through the constructed multi-scale multi-kernel learning, feature extraction of the data is realized, and classification and identification of ship targets by the radar are completed.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology. Background Technology

[0002] High-resolution one-dimensional range profiles (HRRPs) can reflect detailed physical structural features of targets, such as scatterer distribution and target size. Compared with SAR and ISAR images, HRRPs are simpler to implement and easier to process, and are currently widely used in Radar Automatic Target Recognition (RATR). HRRP data is characterized by high dimensionality and nonlinearity. Kernel methods can map nonlinearly inseparable data to a high-dimensional feature space, making the data linearly separable, thus solving complex nonlinear problems using linear methods. While kernel methods can effectively solve nonlinear classification problems, kernel learning typically uses a single kernel function and is highly dependent on the choice of the underlying kernel function. In practical applications, single-kernel learning cannot effectively solve problems when high-dimensional or imbalanced sample sets are encountered.

[0003] To better address the nonlinearity of high-resolution one-dimensional range images, multi-kernel learning methods are commonly used. Multi-kernel learning combines the learning capabilities of multiple single-kernel functions, and its learning effect can be flexibly adjusted based on the weights of different kernel functions, adapting to different datasets. Multi-kernel learning can effectively solve nonlinear problems in recognition. For example, in 2018, the doctoral dissertation from Xi'an University of Electronic Science and Technology, "A Radar Target Recognition Method Based on Multi-kernel and Multi-task Learning," used a data-based adaptive kernel function as the basic kernel function and constructed an adaptive weighted multi-kernel function based on kernel alignment, effectively improving the model's flexibility and applying multi-kernel learning to SAR image target recognition. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-scale, multi-kernel learning target recognition method based on one-dimensional radar range profiles, effectively achieving refined classification and recognition of surface ship targets. The technical solution includes:

[0005] Step (1) Perform multi-scale transformation on the one-dimensional distance image training samples to form multi-scale transformation features at different scales;

[0006] Step (2) Perform kernel space transformation on each multi-scale transformation feature to form kernel space transformations under different kernel functions;

[0007] Step (3) Perform linear synthesis of each kernel space transformation to construct a multi-scale multi-kernel learning projection matrix;

[0008] Step (4) The obtained training sample projection vectors are fed into the SVM classifier for training;

[0009] Step (5) Use the constructed multi-scale multi-kernel learning projection matrix to obtain the projection vector of the test sample;

[0010] Step (6) sends the projection vector of the obtained test sample into the trained SVM classifier for ship target classification and recognition.

[0011] The step (1) further includes:

[0012] Step A: Construct a multi-scale transformation model and set the range and span of the multi-scale transformation parameters;

[0013] Step B involves constructing a discrete KL divergence model and calculating the number of multi-scale transformations.

[0014] Step C performs multi-scale transformation on the one-dimensional distance image training samples to form multi-scale transformation features;

[0015] The step (2) further includes:

[0016] Step D involves constructing a basic kernel function model using the contract scalar method;

[0017] Step E performs kernel space transformation on each multi-scale transformation feature to obtain kernel space transformations under different basic kernel functions;

[0018] The step (3) further includes:

[0019] Step F utilizes a multi-scale, multi-kernel learning method to construct a synthetic kernel function model;

[0020] Step G involves calculating the feature projection matrix using kernel principal component analysis.

[0021] Compared with the prior art, the significant advantages of this invention are:

[0022] The method of calculating the number of multi-scale transformations using the discrete KL divergence model can quickly and effectively determine the number of multi-scale transformations, and this method has the advantages of low computational cost and high operating efficiency. The method of constructing a basic kernel function model using the congruent scaling method can accurately and effectively eliminate the influence of pose sensitivity of one-dimensional range image samples with different poses. The method of constructing a synthetic kernel function model using a multi-scale multi-kernel learning method makes full use of the sample data information. While constructing a multi-scale multi-kernel learning architecture relationship of the data to the greatest extent, it can extract richer low-dimensional features for recognition, making it easier to represent the differences between different samples and effectively classify similar samples. The proposed invention has high application value in the field of radar target recognition.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0024] Figure 1 This is a data processing flowchart of the present invention. Detailed Implementation

[0025] This invention proposes a radar one-dimensional range profile multi-scale multi-kernel learning target recognition method. This method maximizes the preservation of the multi-scale multi-kernel learning architecture relationships in the data and maximizes class separability in the reduced-dimensional Hilbert space, extracting richer low-dimensional features to achieve refined classification and recognition of surface ships by radar. See the attached flowchart for the invention. Figure 1 The preferred implementation steps are as follows:

[0026] Step (1): Perform multi-scale transformation on the one-dimensional distance image training samples to form multi-scale transformation features at different scales. The method is as follows:

[0027] Step A: Construct a multi-scale transformation model, and set the range and span of the multi-scale transformation parameters, specifically as follows:

[0028] Given a data sample set X = [X1, X2, ..., X...] C ]∈R n×N In this context, C represents the total number of sample categories, N represents the total number of samples, and each category contains N samples. k One sample and It is the sample set of the k-th type of target, in which It is the lth of the kth type of target k There are n samples, where n is the dimension of the sample data. One-dimensional distance image samples. The multiscale transformation is expressed as

[0029]

[0030] Midterm This represents the convolution operation, where u represents a one-dimensional distance image unit, and LOG(u, σ) represents the distance unit. s Let be the Laplace kernel of Gauss, and its expression is:

[0031]

[0032] σ s Let σ represent the multiscale coefficients, s = 1, 2, 3, ..., S, where S represents the total number of multiscale coefficients. s ∈[0.1, 10], with a span of 0.5.

[0033] Step B: Construct a discrete KL divergence model and calculate the number of multi-scale transformations, specifically:

[0034] (1) Calculate the discrete KL divergence at each scale. Where Z0 is a one-dimensional distance image training sample, Z s The multi-scale transformation features are defined by a scale factor s, where x is the amplitude of a one-dimensional range image range cell, and D is the range factor. s Let be the discrete KL divergence under the scale factor s.

[0035] (2) Calculate the mean of discrete KL divergence. Where S represents the number of multi-scale transformations.

[0036] (3) Calculate the discrete KL divergence D for each KL divergence. s (Z s Z0) and the mean of discrete KL divergence difference Select μ s The minimum number of multiscale transformations corresponds to the optimal number of multiscale transformations, M.

[0037] Step C: Perform multi-scale transformation on the one-dimensional distance image training samples to form multi-scale transformation features, specifically:

[0038] Calculate the k-th type of target at scale σ m The sample set below is represented as:

[0039]

[0040] Furthermore, the multi-scale transformation sample set of one-dimensional distance images is as follows:

[0041]

[0042] Where, σ m Let m represent the m-th multiscale coefficient, where m = 1, 2, 3, ..., M, and M represents the number of optimal multiscale transformations.

[0043] Step (2): Perform kernel space transformation on each multi-scale transformation feature to form kernel space transformations under different kernel functions, specifically:

[0044] Step D: Construct the basic kernel function model using the contract scalar method, specifically as follows:

[0045] Constructing the basic kernel function for multi-core learning:

[0046]

[0047] Where K m Let m be the m-th fundamental kernel function, m = 1, 2, ..., M, where M represents the number of optimal multiscale transformations, x and y represent the distance cells of one-dimensional range image samples, and δ is the fundamental kernel parameter. m =2 -m .

[0048] Step E: Perform kernel space transformation processing on each multi-scale transformation feature to obtain the kernel space transformation under different basic kernel functions, specifically:

[0049] Through the mapping function, Mapped into a high-dimensional feature space H, its representation is as follows:

[0050]

[0051] in σ is the m-th multiscale transformation factor m The corresponding m-th basic kernel function K m .

[0052] Step (3): Perform linear synthesis of the spatial transformations of each kernel to construct a multi-scale, multi-kernel learning projection matrix, specifically as follows:

[0053] Step F: Construct a synthetic kernel function model using a multi-scale, multi-kernel learning method, specifically as follows:

[0054] Constructing a synthetic kernel function model:

[0055]

[0056] Where Θ = [Θ (1) β, Θ (2) β, ..., Θ (M) β]∈R M×M For the synthesis kernel function model:

[0057]

[0058] β = [β1, β2, ..., β] M ] T ∈R M

[0059] M represents the optimal number of multiscale transformations, K m This represents the m-th basic kernel function. For multi-scale, multi-kernel learning coefficients, where σ m is the scale factor for the m-th multiscale transformation, and N represents the length of the one-dimensional range image sample.

[0060] Step G: Calculate the feature projection matrix using kernel principal component analysis, specifically as follows:

[0061] (1) Center the synthesized kernel function:

[0062] Θ′=Θ-1 n Θ-Θ1 n +1 n Θ1 n

[0063] Among them 1 n It is an n×n matrix where all elements have a value of 1 / n.

[0064] (2) Perform eigenvalue decomposition on Θ′ to obtain the eigenvectors corresponding to the first k eigenvalues, and construct the multi-scale multi-kernel learning feature projection matrix v.

[0065] Step (4): The obtained training sample projection vectors are fed into the SVM classifier for training, specifically as follows:

[0066] Obtain the projection vector of the training samples:

[0067] y = v T X

[0068] Where X is the training sample set. This set is then fed into an SVM classifier for classification training.

[0069] Step (5): Using the constructed multi-scale multi-kernel learning projection matrix, obtain the projection vector of the test sample, as follows:

[0070] y′=v T X′

[0071] Where X′ is the training sample set.

[0072] Step (6): Input the obtained test sample projection vector into the trained SVM classifier for ship target classification and recognition.

Claims

1. A target recognition method based on multi-scale, multi-kernel learning of one-dimensional radar range profiles, characterized in that: Step (1): Perform multi-scale transformation on the one-dimensional distance image training samples to form multi-scale transformation features at different scales; Step (2): Perform kernel space transformation on each multi-scale transformation feature to form kernel space transformations under different kernel functions; Step (3): Perform linear synthesis of each kernel space transformation to construct a multi-scale multi-kernel learning projection matrix; Step (4): Input the obtained training sample projection vectors into the SVM classifier for training; Step (5): Use the constructed multi-scale multi-kernel learning projection matrix to obtain the projection vector of the test sample; Step (6): Input the projection vector of the obtained test sample into the trained SVM classifier for ship target classification and recognition. The step (1) further includes: Step A: Construct a multi-scale transformation model and set the range and span of the multi-scale transformation parameters; Step B: Construct a discrete KL divergence model and calculate the number of multi-scale transformations. First, calculate the discrete KL divergence at each scale. ； in Z 0 represents a one-dimensional distance image training sample. Z s scale factor s Multi-scale transformation characteristics under the following conditions x For the amplitude of a one-dimensional distance image unit, D s scale factor s The discrete KL divergence is calculated, and then the mean of the discrete KL divergence is calculated. ； in S This indicates the number of multi-scale transformations, and finally, the discrete KL divergence is calculated. Mean of discrete KL divergence The difference: ； Select The number of multiscale transformations corresponding to the minimum is the optimal number of multiscale transformations; Step C: Perform multi-scale transformation on the one-dimensional distance image training samples to form multi-scale transformation features; The step (2) further includes: Step D: Construct the basic kernel function model using the contract scalar method; Step E: Perform kernel space transformation on each multi-scale transformation feature to obtain kernel space transformations under different basic kernel functions; The step (3) further includes: Step F: Construct a synthetic kernel function model using a multi-scale, multi-kernel learning method; Step G: Calculate the feature projection matrix using kernel principal component analysis.

2. The radar one-dimensional range image multi-scale multi-kernel learning target recognition method according to claim 1, characterized in that: Step D further includes: ； in K m Indicates the first m One basic kernel function , M This represents the optimal number of multiscale transformations. x, y Represents the distance cell of a one-dimensional distance image sample, with basic kernel parameters. .

3. The radar one-dimensional range image multi-scale multi-kernel learning target recognition method according to claim 1, characterized in that: Step F includes: ； in Synthetic kernel function model, M This represents the optimal number of multiscale transformations. K m Indicates the first m One basic kernel function These are multi-scale, multi-kernel learning coefficients, where For the first m The scaling factor of a multi-scale transformation.

Images