CN103941244B - A kind of radar target-range image local optimum subspace identification method - Google Patents

Abstract

The invention provides a kind of radar target-range image local optimum subspace identification method, effectively improve the recognition performance to radar target, the method calculates nearest inter-object distance scatter matrix and nearest between class distance scatter matrix first with training data, then, a local optimum subspace is set up by optimum ratio criterion, extract target characteristic, use minimum distance classifier to classify, finally determine input classification belonging to target；Specially utilize radar target-range image trained vector x_ijDetermine vectorAnd vectorDetermine vectorWithDetermine matrix D_WAnd D_B, determine m vector a of local optimum subspace₁,a₂,…,a_m, according to λ_iWith vector a_i(i=1,2 ..., m) determine local optimum subspaceUtilize trained vector projection in the A of subspace to determine template base, determine input target one-dimensional range profile x_tLocal optimum sub-picture, determine the distance between local optimum sub-picture and library template vector, utilize minimum distance classifier to determine input classification belonging to target one-dimensional range profile.

Description

A kind of radar target-range image local optimum subspace identification method

Technical field

The invention belongs to Technology of Radar Target Identification field, relate to a kind of radar target-range image local optimum subspace and identify Method.

Background technology

Subspace method is classical mode identification method, is widely used in image recognition, recognition of face, knows at radar target Also a lot of application is had in not.Common proper subspace method and Canonical sub-space method all achieve well in radar target recognition Recognition effect.Wherein, proper subspace can keep the energy of initial data in low dimensional feature space, but from classification performance Say it is not optimum.For proper subspace, Canonical sub-space is by making inter-object distance minimum and between class distance maximum Extract target characteristic, improve target identification performance to a certain extent.

But, Canonical sub-space method uses average and average distribution between class structure in macroscopically describing class in class between class, but from office Seeing in portion and be not likely to be optimum, Canonical sub-space dimension is determined by target classification number simultaneously, and target classification a few hours can cause point The loss of category information.Therefore, the recognition performance of Canonical sub-space method has further room for improvement.

Summary of the invention

Object of the present invention is to provide a kind of radar target-range image local optimum subspace identification method, effectively carry The high recognition performance to radar target, its technical scheme is:

A kind of radar target-range image local optimum subspace identification method, it is characterised in that first with training data meter Calculate nearest inter-object distance scatter matrix and nearest between class distance scatter matrix, then, set up a local by optimum ratio criterion Optimal subspace, extracts target characteristic, uses minimum distance classifier to classify, and finally determines input classification belonging to target, Comprise the following steps:

1) radar target-range image trained vector x is utilized_ijDetermine vectorAnd vector

2) according to x_ij、WithDetermine vectorWith

3) according to vectorWithDetermine matrix D_WAnd D_B；

4) matrix D is utilized_WAnd D_BDetermine m vector a of local optimum subspace₁,a₂,…,a_m；

5) according to λ_iWith vector a_i(i=1,2 ..., m) determine local optimum subspace

6) trained vector projection in the A of subspace is utilized to determine template base；

7) input target one-dimensional range profile x is determined_tLocal optimum sub-picture；

8) distance between local optimum sub-picture and library template vector is determined；

9) minimum distance classifier is utilized to determine input classification belonging to target one-dimensional range profile.

Further preferably, utilize the local optimum sub-picture of every classification target one-dimensional range profile training sample to set up template base, determine office Distance between portion's optimum picture and library template, judges target classification by minimum distance criterion.

Further preferably, it is determined that the concrete way of target is: to matrixCarry out eigen decomposition, by front m maximum intrinsic Value and corresponding eigenvector determine local optimum subspaceTraining objective one-dimensional range profile By formula y=A^TX calculates local optimum sub-picture, and the local optimum sub-picture of every classification target training one-dimensional range profile is formed corresponding mesh Target template base { y_ij, wherein y_ijIt it is the i-th classification target jth training one-dimensional range profile vector；To input target one-dimensional away from From as by x_t, y=A^TX calculates local optimum as y_t, and calculate following distance:

s_ij=| | y_t-y_ij| | i=1,2 ... g；J=1,2 ..., N_i

Determine the minimum range in the i-th class target:

s_i=min{si_j}

If

$k=\arg \underset{\left\{i\right\}}{\min }\left\{s_i\right\},$ Then sentencing input target is kth class.

Beneficial effects of the present invention:

The present invention utilizes training sample data to set up local optimum subspace to extract target characteristic, makes one-dimensional range profile correspondence special The nearest inter-object distance levied minimizes and nearest between class distance reaches maximum, improves target classification performance；Meanwhile, local optimum The size of subspace is unrelated with number of targets, it is possible to obtain the subspace of suitable size, more fully keeps the classification of initial data to believe Breath, is effectively improved the radar recognition performance to multi-class targets.

Accompanying drawing explanation

Fig. 1 is the flow chart of radar target-range image local optimum subspace identification method of the present invention.

Detailed description of the invention

Below in conjunction with detailed description of the invention and accompanying drawing, the present invention is described in further details.

Local optimum varitron space:

If x_ij(n ties up column vector) is i-th^thClassification target jth^thIndividual training one-dimensional range profile, i=1,2 ..., g；J=1,2 ..., N_i, N₁+N₂+…+N_g=N, wherein N_iIt is i-th^thClassification target training one-dimensional range profile sample number, N is training one-dimensional distance As total sample number.Definition x_ijNearest class in the outer one-dimensional range profile sample of one-dimensional range profile sample and nearest class be

$x_{\mathrm{ij}}^W=\underset{\left\{x_{\mathrm{ir}}\right\}}{\arg \min }{\left|\right|x_{\mathrm{ij}}-x_{\mathrm{ir}}\left|\right|}_p,r=\mathrm{1,2},...,N_i---\left(1\right)$

$x_{\mathrm{ij}}^B=\underset{\left\{x_{\mathrm{kr}}\right\}}{\arg \min }{\left|\right|x_{\mathrm{ij}}-x_{\mathrm{kr}}\left|\right|}_p,k=\mathrm{1,2},...,g,k\≠i,r=\mathrm{1,2},...,N_k---\left(2\right)$

Wherein | | | |_pFor p norm,For x_ijIn corresponding nearest class one-dimensional decent,For x_ijCorresponding nearest class is outer one-dimensional decent This.Then sample x_ijCorresponding nearest inter-object distance vector nearest between class distance vector is

$d_{\mathrm{ij}}^W=|x_{\mathrm{ij}}-x_{\mathrm{ij}}^W|---\left(3\right)$

$d_{\mathrm{ij}}^B=|x_{\mathrm{ij}}-x_{\mathrm{ij}}^B|---\left(4\right)$

Wherein | | represent that vector element takes absolute value,WithIt is respectively x_ijThe nearest nearest between class distance of inter-object distance vector vow Amount.

Making a is that arbitrary n ties up column vector, calculates following ratio

$F=\frac{\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{a}^T{d}_{\mathrm{ij}}^W{\left(d_{\mathrm{ij}}^W\right)}^{T}a}{\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{a}^T{d}_{\mathrm{ij}}^B{\left(d_{\mathrm{ij}}^B\right)}^{T}a}=\frac{a^T{D}_{W}a}{a^T{D}_{B}a}---\left(5\right)$

Wherein D_WFor the nearest inter-object distance scatter matrix of each training sample, D_BNearest between class distance for each training sample spreads square Battle array, its calculation expression is as follows

$D_W=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{d}_{\mathrm{ij}}^W{\left(d_{\mathrm{ij}}^W\right)}^T---\left(6\right)$

$D_B=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{d}_{\mathrm{ij}}^B{\left(d_{\mathrm{ij}}^B\right)}^T---\left(7\right)$

Formula (5) shows, ratio F is the least, and the nearest inter-object distance of each sample is the least, and the nearest class distance of each sample is the biggest, from And distribution between class structure in the premium class of each sample can be obtained, be conducive to strengthening classification performance.Formula (5) both sides to vector a derivation, and , in zero, abbreviation can obtain to make result wait until

$D_B^{-1}{D}_{W}a=\mathrm{\λa}---\left(8\right)$

Wherein λ is characteristic value, and a is the eigenvector that λ is corresponding.If the finite eigenvalues in formula (8) put in order for λ₁＞ λ₂…λ_m... ＞ 0, corresponding eigenvector is a₁,a₂,…,a_m..., take front m maximum finite eigenvalues and the basis of correspondence Levy vector and form following matrix

$A=[\sqrt{{\mathrm{\λ}}_1}{a}_1\sqrt{{\mathrm{\λ}}_2}{a}_2...\sqrt{{\mathrm{\λ}}_m}{a}_m]---\left(9\right)$

Then matrix A is called optimal partial subspace.Sample x_ijTo subspace, A projects

y_ij=A^Tx_ij (10)

Then claim yi_jFor xi_jCorresponding projection vector.Then, projection vector is utilized to be identified as target characteristic.

Radar target recognition based on local optimum subspace:

Assume there is g class target, every classification target one-dimensional range profile training sample set can be utilized to set up local by formula (8) and formula (9) Optimal subspace, the one-dimensional range profile of training objective obtains sub-picture by formula (10) to local optimum subspace projection, by every class target The local optimum sub-picture of one-dimensional range profile training sample form the library template vector of such target, then the i-th classification target template base is

$\{y_{i1}{y}_{i2}...y_{{\mathrm{iN}}_i}\}---\left(11\right)$ One-dimensional range profile x to input target_t, can calculate optimum Nonlinear Orthogonal projection vector by formula (10) is y_t, and calculate with

Lower distance

s_ij=| | y_t-y_ij| | i=1,2 ... g；J=1,2 ..., N_i (12)

$s_i=\underset{\left\{j\right\}}{\min }\left\{s_{\mathrm{ij}}\right\}---\left(13\right)$

If

$k=\arg \underset{\left\{i\right\}}{\min }\left\{s_i\right\}---\left(14\right)$

Then sentencing input target is kth class.

Accompanying drawing illustrates the flow chart of the radar target-range image local optimum subspace identification method of the present invention.Flow chart is opened Start from step 201.

In step 202, g classification target is utilized to train one-dimensional range profile vector x_ijIt is defined below vector:

$x_{\mathrm{ij}}^W=\underset{\left\{x_{\mathrm{ir}}\right\}}{\arg \min }{\left|\right|x_{\mathrm{ij}}-x_{\mathrm{ir}}\left|\right|}_p$

$x_{\mathrm{ij}}^B=\underset{\left\{x_{\mathrm{kr}}\right\}}{\arg \min }{\left|\right|x_{\mathrm{ij}}-x_{\mathrm{kr}}\left|\right|}_p$

Wherein r=1,2 ..., N_i；K=1,2 ..., g, k ≠ i, r=1,2 ..., N_k；G is target classification number, N_iIt it is the i-th classification target Number of training；

In step 2031, determine vector:

$d_{\mathrm{ij}}^W=|x_{\mathrm{ij}}-x_{\mathrm{ij}}^W|$

$d_{\mathrm{ij}}^B=|x_{\mathrm{ij}}-x_{\mathrm{ij}}^B|$

In step 2032, determine matrix:

$D_W=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{d}_{\mathrm{ij}}^W{\left(d_{\mathrm{ij}}^W\right)}^T$

$D_B=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{d}_{\mathrm{ij}}^B{\left(d_{\mathrm{ij}}^B\right)}^T$

In step 2033, determine the vector in local optimum subspace: a_iFor matrixBasis corresponding to m dominant eigenvalue Levy vector；

In step 2034, determine local optimum subspace:Wherein λ_iFor vector a_iCorresponding intrinsic Value；

In step 204, determine every class target template base:

In step 2051, determine input target one-dimensional range profile x_tLocal optimum sub-picture: y_t=A^Tx_t；

In step 2052, determine the distance between local optimum sub-picture and To Template vector:

s_ij=| | y_t-y_ij| | i=1,2 ... g；J=1,2 ..., N_i

$s_i=\underset{\left\{j\right\}}{\min }\left\{s_{\mathrm{ij}}\right\}$

In step 2053, determine input classification belonging to target one-dimensional range profile:

The flow process of the radar target-range image local optimum subspace identification method according to the present invention ends at step 206.

The simulation experiment result: the effectiveness of extracting method in order to verify, carries out following emulation experiment.

Design four kinds of point targets: " | " font, " V " font, " doing " font and " little " font target.Radar emission arteries and veins The a width of 150MHZ of band (range resolution ratio is 1m, and radar radially sampling interval is 0.5m) of punching, goal setting is homogenous diffusion Point target, " | " scattering point of target is 5, the scattering of its excess-three target is counted and is 9.It it is 0 ° at object attitude angle ～in the range of 50 ° in the one-dimensional range profile of 1 °, take object attitude angle be 0 °, 2 °, 4 °, 6 ° ..., 50 ° One-dimensional range profile be trained, the one-dimensional range profile of remaining attitude angle is as test data, then every classification indicates 25 surveys Sample is originally.

To four kinds of targets (" | " font target, " V " font target, " doing " font target and " little " font target), In the range of attitude angle 0o～50o, utilize radar target-range image local optimum subspace identification method herein and Canonical sub-space method has carried out identifying experiment, and result is as shown in Table 1.

From table one, to target " | ", the discrimination of Canonical sub-space method is 88%, and radar target herein away from From being 96% as the discrimination of local optimum subspace identification method；To target " V ", identification based on Canonical sub-space method Rate is 86%, and the discrimination of radar target-range image local optimum subspace identification method herein is 90%；To mesh Mark " doing ", discrimination based on Canonical sub-space method is 78%, and radar target-range image local optimum herein The discrimination of space recognition method is 85%；To target " little ", the discrimination of Canonical sub-space method is 85%, and herein The discrimination of radar target-range image local optimum subspace identification method is 90%.On average, to four class targets, The correct recognition rata of radar target-range image local optimum subspace identification method herein is higher than Canonical sub-space method, Illustrate that radar target-range image local optimum subspace identification method herein can improve the identity of multi-class targets really Energy.

The recognition result of one or two kind of method of table

Claims (3)

1. a radar target-range image local optimum subspace identification method, it is characterised in that first with training data Calculate nearest inter-object distance scatter matrix and nearest between class distance scatter matrix, then, set up an office by optimum ratio criterion Portion's optimal subspace, extracts target characteristic, uses minimum distance classifier to classify, and finally determines input class belonging to target , do not comprise the following steps:

1) radar target-range image trained vector x is utilized_ijDetermine vectorAnd vectorDescribed vectorFor x_ijCorresponding Nearest class in one-dimensional decent, vectorFor x_ijOuter one-dimensional decent of corresponding nearest class；

2) according to x_ij、WithDetermine vectorWithDescribed vectorFor x_ijCorresponding nearest inter-object distance vector, VectorFor x_ijCorresponding nearest between class distance vector；

3) according to vectorWithDetermine matrix D_WAnd D_B:

$D_W=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{d}_{ij}^W{\left(d_{ij}^W\right)}^T,D_B=\underset{i=1}{\overset{g}{\mathrm{\Σ}}}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}{d}_{ij}^B{\left(d_{ij}^B\right)}^T;$

Wherein, g is target classification number, N_iIt is the i-th classification target number of training, described matrix D_WFor each training sample Nearest inter-object distance scatter matrix, matrix D_BNearest between class distance scatter matrix for each training sample；

4) matrix D is utilized_WAnd D_BDetermine m vector a of local optimum subspace₁,a₂,…,a_m；

5) according to λ_iWith vector a_i, i=1,2 ..., m determines local optimum subspaceDescribed λ_iRepresent Vector a_i, i=1,2 ..., the characteristic value that m is corresponding；

6) trained vector projection in the A of subspace is utilized to determine template base；

7) input target one-dimensional range profile x is determined_tLocal optimum sub-picture；

8) distance between local optimum sub-picture and library template vector is determined；

9) minimum distance classifier is utilized to determine input classification belonging to target one-dimensional range profile.

2. a kind of radar target-range image local optimum subspace identification method as described in claim 1, it is characterised in that Template base set up by the local optimum sub-picture utilizing every classification target one-dimensional range profile training sample, determines local optimum sub-picture and storehouse mould Distance between plate, judges target classification by minimum distance criterion.

3. a kind of radar target-range image local optimum subspace identification method as described in claim 1, it is characterised in that The concrete way of judgement target is: to matrixCarry out eigen decomposition, by front m dominant eigenvalue and corresponding intrinsic to Amount determines local optimum subspaceTraining objective one-dimensional range profile presses formula y=A^TX calculates Local optimum sub-picture, x are training objective one-dimensional range profile, by local optimum of every classification target training one-dimensional range profile as group Become the template base { y of respective objects_ij, wherein y_ijIt it is the i-th classification target jth training one-dimensional range profile vector；To input target One-dimensional range profile press x_t, y=A^TX calculates local optimum as y_t, and calculate following distance:

s_ij=| | y_t-y_ij| | i=1,2 ... g；J=1,2 ..., N_i

Determine the minimum range in the i-th class target:

s_i=min{s_ij}

If

Then sentencing input target is kth class.