CN113267757A - Unknown target discrimination method based on K times of singular value decomposition dictionary learning - Google Patents

Abstract

The invention belongs to the technical field of target recognition, and particularly relates to an unknown target discrimination method based on K-time singular value decomposition dictionary learning. According to the method, firstly, dictionary learning is carried out on a training one-dimensional range profile sample set by using a K-time singular value decomposition method to obtain an over-complete dictionary, and then a reconstruction error is obtained through sparse representation based on the over-complete dictionary to realize the discrimination of an unknown target. The sparsity of the one-dimensional range profile is effectively described through sparse representation, and the reconstruction capability of dictionary atoms on the one-dimensional range profile of the target can be enhanced through the K-time singular value decomposition dictionary learning algorithm, so that the discrimination performance of the unknown target is improved.

Description

Unknown target discrimination method based on K times of singular value decomposition dictionary learning

Technical Field

The invention belongs to the technical field of target recognition, and particularly relates to an unknown target discrimination method based on K-time singular value decomposition dictionary learning.

Background

The one-dimensional range image is the coherent sum of echo signals returned by the scattering points of the target on each range unit, provides distribution information of the scattering points of the target along the range direction and structural information of the target, and is beneficial to classification and identification of radar targets. The conventional radar target one-dimensional range profile identification method needs to acquire target data in advance, train and establish a template library to realize correct identification. However, in practical applications, data of some targets cannot be acquired in advance, so that the targets cannot participate in training and are mistakenly recognized as known targets. Therefore, it is necessary to discriminate the unknown object before the recognition.

Currently, radar unknown target discrimination is mainly divided into a discrimination method based on data generation and a threshold discrimination method. The discrimination method based on data generation generates data through a statistical model, and discriminates a target by using two discrimination methods, but the generated data and the actually measured data have deviation and cannot meet the actual requirement. The discrimination method based on the threshold extracts target features through an algorithm to construct a discrimination threshold so as to realize discrimination of the target, such as a feature subspace method, support vector data domain description, a neural network and the like. The characteristic subspace method is used for identifying the sub-image obtained through projection, the support vector data domain description method is used as a single classification algorithm and needs to optimize multiple parameters, the neural network method needs a large amount of data for training the network and multiple training parameters, the methods need a large amount of data for training, but sample data obtained in practical application is limited, and the identification performance of the existing method is reduced. Therefore, the conventional method has room for further improvement.

Disclosure of Invention

Aiming at the problems, the invention provides an unknown target discrimination method based on K times of singular value decomposition dictionary learning. The method includes the steps that an over-complete dictionary is obtained through a K-time singular value decomposition dictionary learning algorithm, and then sparse reconstruction errors of one-dimensional range profiles of targets to be recognized are calculated based on the dictionary, so that the unknown targets are distinguished. The sparsity of the one-dimensional range profile is effectively described through sparse representation, and the reconstruction capability of dictionary atoms on the target one-dimensional range profile can be enhanced through the K-time singular value decomposition dictionary learning algorithm, so that the discrimination performance is improved.

The technical scheme of the invention is as follows:

an unknown target discrimination method based on K times of singular value decomposition dictionary learning comprises the following steps:

s1, setting a known target one-dimensional range profile training sample set Y as:

Y＝[y₁,y₂,...,y_N]

wherein, y_iIs the ith m-dimensional training one-dimensional range profile sample, and N is the number of training samples;

performing sparse representation on the training sample set Y by using the overcomplete dictionary D:

Y＝DX

D＝[d₁,d₂,...,d_K]

wherein d is_kThe kth atom of the dictionary D is represented, K is 1,2, … K, K is the number of dictionary atoms, and X is a K × N-dimensional sparse matrix:

X＝[x₁,x₂,...,x_N]

wherein x is_iDenotes y_iA corresponding sparse vector;

s2, obtaining an over-complete dictionary based on K times of singular value decomposition dictionary learning, and the specific method comprises the following steps:

s21, inputting: knowing a target training set Y, an initial dictionary D and sparsity T, presetting the iteration number as T, and simultaneously setting a parameter q as 1;

s22, sparse representation: calculating a sparse matrix X in Y ═ DX by using an OMP (orthogonal matching pursuit) algorithm;

s23, dictionary updating: updating dictionary atom d one by one_kK is 1, 2.. K, which specifically includes:

s231, calculating an error matrix

Wherein x^jRepresents the jth row in the matrix X;

s232, from x^kThe indices of the medium non-zero elements constitute an index vector:

w_k＝{i|1≤i≤N,x^k(i)≠0}

wherein x is^k(i) Denotes x^kThe ith element of (1), from w_k(K1, 2, … K) constitutes an index matrix Ω_k，Ω_kMiddle (w)_kRow and column (p) has an element value of 1, p is 1,2, … | w_kI, other element values are 0, | w_kL is w_kThe number of middle elements;

s233, order

S234, pair

Singular value decomposition is carried out:

wherein U is [ U ]₁,u₂,…,u_m]Is formed by singular vectors u₁、u₂、…u_mThe left singular matrix of the composition is,

is formed by singular vectors

Right singular matrix of components, | w_k| represents w_kThe number of the middle elements, sigma is a diagonal matrix formed by singular values;

s235, updating dictionary atoms: d_k＝u₁；

S236, update

λ₁Is the largest singular value in the singular value matrix Σ;

s24, making q equal to q +1, if q > t, stopping iteration to obtain an overcomplete dictionary D, otherwise, returning to step S22;

s3, setting the one-dimensional range profile verification set of the known target as Y_v＝[y_v,1,y_v,2,...,y_v,L]Wherein, y_v,lRepresenting the L-th verification one-dimensional range profile, L representing the number of verification one-dimensional range profiles, and using the obtained overcomplete dictionary D to y_v,lPerforming sparse representation to obtain a sparse vector x_v,lThen the reconstruction error is:

e_v,l＝||y_v,l-Dx_v,l||₂l＝1,2,...,L

wherein | · | purple sweet₂Denotes a norm of order 2, e_v,lIs y_v,lCorresponding sparse reconstruction errors, forming an error sequence e from the reconstruction errors of the one-dimensional range profile in the verification set_v,1、e_v,2、…e_v,LIf yes, the decision threshold τ is:

τ＝min(e_v,1,e_v,2,…e_v,L)+γ[max(e_v,1,e_v,2,…e_v,L)-min(e_v,1,e_v,2,…e_v,L)]

0≤γ≤1

wherein gamma is a coefficient, and max (-) and min (-) take the maximum value and the minimum value in the sparse reconstruction error sequence respectively;

s4, performing sparse representation on the one-dimensional range profile of the target to be recognized by using the over-complete dictionary D to obtain a reconstruction error, and comparing the reconstruction error with a discrimination threshold to determine a discrimination result; setting a one-dimensional range profile data set of a target to be identified as Y_t＝[y_t,1,y_t,2,...,y_t,M]Wherein y is_t,hFor the h-th test one-dimensional range profile sample, M represents the number of the tested one-dimensional range profiles, and the sparse vector x is calculated through an OMP algorithm_t,hH is 1,2, … M, and one-dimensional range image y of the target to be recognized is calculated_t,hThe sparse reconstruction error on the overcomplete dictionary D is:

e_t,h＝||y_t,h-Dx_t,h||₂

will reconstruct the error e_t,hComparing with a discrimination threshold tau, and if the reconstruction error e of the one-dimensional range profile of the target to be recognized is the same as the one-dimensional range profile of the target to be recognized_t,hIf the difference is less than the discrimination threshold tau, the target is discriminated as a known target, otherwise, the target is discriminated as an unknown target.

The invention has the beneficial effects that: the K-time singular value decomposition dictionary learning algorithm can enhance the reconstruction capability of dictionary atoms on the one-dimensional range profile of the target, so that the discrimination performance of the unknown target is improved.

Detailed Description

The practical applicability of the present invention will be described below with reference to simulation experiments.

The data used in the simulation experiment are one-dimensional range profile data of five types of airplanes obtained through electromagnetic simulation calculation software, wherein the five types of airplanes are Ah64, An26, F15, B1B and B52. Wherein, the radar carrier frequency is 6GHz, and the bandwidth is 400 MHz. The pitch angle of each type of airplane is 3 degrees, the azimuth angle range is 0-180 degrees, one-dimensional distance image is collected every 0.1 degree, each type of airplane is provided with 1801 one-dimensional distance images, and each one-dimensional distance image comprises 320 distance units.

300 pieces of one-dimensional range profile data of each type of known target in the range of 0-60 degrees of azimuth angle are selected at intervals of 0.2 degrees to form a training target set, the rest one-dimensional range profile samples are used as a verification target set, and the one-dimensional range profile data of each type of unknown target in the same range of azimuth angle as the known target is selected at the same intervals to be used as a test target set. The three data sets were amplitude normalized.

The number of dictionary atoms K is 400, the sparse vector sparsity T is 20, and the coefficient gamma is set to be 0.15.

Two types of airplanes are selected as known targets, the other three types of airplanes are selected as unknown targets, the unknown targets are judged by using a K-time singular value decomposition dictionary learning method, and an average correct judgment rate of 88% is obtained, so that the effectiveness of the method is verified.

Claims (1)

1. An unknown target discrimination method based on K times of singular value decomposition dictionary learning is characterized by comprising the following steps:

s1, setting a known target one-dimensional range profile training sample set Y as:

Y＝[y₁,y₂,...,y_N]

wherein, y_iIs the ith m-dimensional training one-dimensional range profile sample, and N is the number of training samples;

performing sparse representation on the training sample set Y by using the overcomplete dictionary D:

Y＝DX

D＝[d₁,d₂,...,d_K]

wherein d is_kThe kth atom of the dictionary D is represented, K is 1,2, … K, K is the number of dictionary atoms, and X is a K × N-dimensional sparse matrix:

X＝[x₁,x₂,...,x_N]

wherein x is_iDenotes y_iA corresponding sparse vector;

s2, obtaining an over-complete dictionary based on K times of singular value decomposition dictionary learning, and the specific method comprises the following steps:

s21, a known target training set Y, an initial dictionary D, sparsity T, and presetting the iteration number as T, and meanwhile, setting a parameter q as 1;

s22, calculating Y ═ DX sparse matrix X using OMP (orthogonal matching pursuit) algorithm;

s23, updating dictionary atom d one by one_kK is 1, 2.. K, which specifically includes:

s231, calculating an error matrix

Wherein x^jRepresents the jth row in the matrix X;

s232, from x^kThe indices of the medium non-zero elements constitute an index vector:

w_k＝{i|1≤i≤N,x^k(i)≠0}

wherein x is^k(i) Denotes x^kThe ith element of (1), from w_k(K1, 2, … K) constitutes an index matrix Ω_k，Ω_kMiddle (w)_kRow and column (p) has an element value of 1, p is 1,2, … | w_kI, other element values are 0, | w_kL is w_kThe number of middle elements;

s233, order

S234, pair

Singular value decomposition is carried out:

wherein U is [ U ]₁,u₂,…,u_m]Is formed by singular vectors u₁、u₂、…u_mLeft singular moment of compositionThe number of the arrays is determined,

is formed by singular vectors

Right singular matrix of components, | w_k| represents w_kThe number of the middle elements, sigma is a diagonal matrix formed by singular values;

s235, updating dictionary atoms: d_k＝u₁；

S236, updating x^k:

λ₁Is the largest singular value in the singular value matrix Σ;

s24, making q equal to q +1, if q > t, stopping iteration to obtain an overcomplete dictionary D, otherwise, returning to step S22;

s3, setting the one-dimensional range profile verification set of the known target as Y_v＝[y_v,1,y_v,2,...,y_v,L]Wherein, y_v,lRepresenting the L-th verification one-dimensional range profile, L representing the number of verification one-dimensional range profiles, and using the obtained overcomplete dictionary D to y_v,lPerforming sparse representation to obtain a sparse vector x_v,lThen the reconstruction error is:

e_v,l＝||y_v,l-Dx_v,l||₂ l＝1,2,...,L

wherein | · | purple sweet₂Denotes a norm of order 2, e_v,lIs y_v,lCorresponding sparse reconstruction errors, forming an error sequence e from the reconstruction errors of the one-dimensional range profile in the verification set_v,1、e_v,2、…e_v,LIf yes, the decision threshold τ is:

τ＝min(e_v,1,e_v,2,…e_v,L)+γ[max(e_v,1,e_v,2,…e_v,L)-min(e_v,1,e_v,2,…e_v,L)]

0≤γ≤1

wherein gamma is a coefficient, and max (-) and min (-) take the maximum value and the minimum value in the sparse reconstruction error sequence respectively;

s4, performing sparse representation on the one-dimensional range profile of the target to be recognized by using the over-complete dictionary D to obtain a reconstruction error, and comparing the reconstruction error with a discrimination threshold to determine a discrimination result; setting a one-dimensional range profile data set of a target to be identified as Y_t＝[y_t,1,y_t,2,...,y_t,M]Wherein y is_t,hFor the h-th test one-dimensional range profile sample, M represents the number of the tested one-dimensional range profiles, and the sparse vector x is calculated through an OMP algorithm_t,hH is 1,2, … M, and one-dimensional range image y of the target to be recognized is calculated_t,hThe sparse reconstruction error on the overcomplete dictionary D is:

e_t,h＝||y_t,h-Dx_t,h||₂

will reconstruct the error e_t,hComparing with a discrimination threshold tau, and if the reconstruction error e of the one-dimensional range profile of the target to be recognized is the same as the one-dimensional range profile of the target to be recognized_t,hIf the difference is less than the discrimination threshold tau, the target is discriminated as a known target, otherwise, the target is discriminated as an unknown target.