CN103217676A - Radar target identification method under noise background based on bispectrum de-noising - Google Patents

Abstract

The invention discloses a radar target identification method under a noise background based on bispectrum de-noising and the method is mainly used for solving the problems in the prior art that target identification is carried out by directly utilizing a distance image which is not de-noised under the noise background so that the identification rate is poor. The realization process comprises the following steps of: normalizing a training sample and extracting the power spectrum characteristic of the normalized training sample; utilizing a power spectrum characteristic training classifier of the training sample to obtain a weight coefficient of the classifier; utilizing a bispectrum de-noising method to carry out de-noising treatment on a testing sample and recover an average distance image which is de-noised, and normalizing the distance image; extracting a power spectrum characteristic of the normalized average distance image; and utilizing the trained classifier to classify the power spectrum characteristic of the normalized de-noised average distance image, and determining a target class mark number. The radar target identification method disclosed by the invention has the advantages of stabilizing noises and recovering the average distance image which is de-noised, and can be used for radar target identification.

Description

Based on radar target identification method under the noise background of two spectrum denoisings

Technical field

The invention belongs to the Radar Technology field, relate to target identification method, can be used for to noise background get off the plane, target such as vehicle discerns.

Background technology

High resolution range profile is the vector sum of target scattering point echo projection on the radar ray direction of wideband radar signal, the general distribution situation of target scattering point echo at range direction can be provided, identification of targets is had important value, thereby become the focus of radar automatic target identification area research.

Owing to the randomness of High Range Resolution, make distance images have the translation sensitive question in the practical application, thereby directly utilize distance images identification need carry out the translation alignment in the appearance of distance window.Distance images alignment schemes commonly used has sliding correlation method and absolute alignment method, and the sliding correlation method precision is higher, but calculation of complex, and definitely alignment method is calculated simply, but precision is lower.Two spectrum signatures have translation invariance, and High Range Resolution can be directly average at two spectral domains, and need not do the translation alignment to it, the computational problem that this has been avoided the distance images alignment to bring.Two spectrum signatures are that zero noise has blind property to having symmetrical probability density function and average arbitrarily, utilize the inhibiting effect of two spectrum signatures to white Gaussian noise, can solve the denoising problem of High Range Resolution under noise background.Two spectrum signatures have kept the whole phase informations of signal except that linear phase, except the position is uncertain, can from two spectrum signatures unique recover the raw range picture.

Two spectrums are inhibited to white noise, but two spectrum denoising of multiple distance images must be considered the first phase sensitive question of multiple distance images.The phase place of multiple distance images comprises by target rotates phase place that causes and the phase place that is caused by the target translation, rotate the phase place that causes by target and comprise identification of targets information, has certain identification value, and the i.e. first phase of distance images again of the phase place that causes by the target translation, under the certain situation of radar signal wavelength, first phase is by the decision of the radial distance of target and radar, does not comprise identification of targets information, therefore identification is not worth.For a C-band radar, if carrier frequency is 6GHz, wavelength is about 5cm, and then the translation of 5cm will make first phase that the variation of 4 π takes place.The very little translation of this explanation will cause that very big first phase changes, and this makes the phase place of multiple distance images be difficult to be utilized, the first phase sensitive question of multiple distance images that Here it is.If can proofread and correct preferably, can eliminate the first phase sensitive question to first phase.

The autofocus algorithm that uses in the existing inverse synthetic aperture radar imaging can be realized the first phase correction of multiple distance images in theory, but be subjected to all multifactor influences such as target range radar distance, the signal to noise ratio (S/N ratio) of target is often lower, if do not remove The noise, autofocus algorithm can not realize under the low signal-to-noise ratio situation that first phase is proofreaied and correct preferably, and then influence the Target Recognition effect.

Summary of the invention

The object of the present invention is to provide a kind ofly, solving above-mentioned prior art mean distance picture after can't recovering denoising under the noise background, and then cause the low deficiency of discrimination based on radar target identification method under the noise background of two spectrum denoisings.

Realize that basic ideas of the present invention are: by using the power spectrum characteristic training linear associated vector machine sorter of training sample, by one group of amplitude is the distance images amplitude, phase place is average two spectrums that the deformation distance picture of distance images phase place twice calculates test sample book, recover mean distance picture after the denoising by average two spectrum, the power spectrum characteristic of mean distance picture after the denoising is input to linearly dependent vector machine sorter, determines the target category label.Concrete steps comprise as follows:

(1) from the radar return database, takes out multi-class targets as training objective, and its echo carried out pulse compression, obtain the distance images sample x of training objective, and decent the x normalization of adjusting the distance, by the weight coefficient W of distance images sample calculation linearly dependent vector machine sorter after the normalization;

(2) radar system as test target, and is carried out pulse compression respectively to R continuous echo of this test target with detected certain unknown object, obtains test target distance images sample set: X={x ₁, x ₂..., x _d..., x _R, d=1 wherein, 2 ..., R, x _dEcho carries out the distance images sample that pulse compression obtains for d time;

(3) mean distance picture after the acquisition denoising:

3a) take out distance images sample in the test target distance images sample set successively, obtain its pair spectrum:

3a1) distance images sample x to taking out _d={ x _d(0), x _d(1) ... x _d(e) ... x _d(N-1)), use following formula to produce deformation distance picture: y _d={ y _d(0), y _d(1) ... y _d(e) ... y _d(N-1) }:

$y_d\left(e\right)=\frac{x_d\left(e\right)\·x_d\left(e\right)}{\left|x_d\left(e\right)\right|},$

X wherein _d(e) be x _dIn e dimension element, e=0,1 ..., N-1, N are the distance images dimension;

y _d(e) be y _dIn e dimension element, e=0,1 ..., N-1, || the mould value is asked in expression;

3a2) decent the x that adjust the distance _dDo Fast Fourier Transform (FFT), obtain x _dFrequency spectrum S _d, to deformation distance as y _dDo Fast Fourier Transform (FFT), obtain y _dFrequency spectrum T _d

3a3) pass through x _dFrequency spectrum S _d, y _dFrequency spectrum T _dObtain x _dTwo spectrum B _d, this pair spectrum B _dBe N dimension square formation, B _dIn p+1 capable, q+1 column element: B _d(p, q)=S _d(p) S _d(q) T _d ^*(p+q), S wherein _d(p) be S _dP dimension element, S _d(q) be S _dQ dimension element, T _d ^*(p+q) be T _dIn the conjugation of p+q dimension element, p=0,1 ..., N-1, q=0,1 ..., N-1;

3b) by step 3a) R that obtains two spectrums, constitute two spectrum signature collection B=(B ₁, B ₂... B _d..., B _R, calculate average two spectrum signature

D=1 wherein, 2 ..., R, B _dBe two spectrum signatures of the d time distance images sample;

3c) by average two spectrum signature B ', obtain the amplitude U and the phase place V of mean distance picture frequency spectrum after the denoising;

3d) amplitude U and the phase place V to mean distance picture frequency spectrum after the denoising carries out inverse fast fourier transform, obtain denoising after mean distance as x ';

(4) mean distance obtains the target category label as x ' by the weight coefficient W of mean distance picture after the normalized denoising and linearly dependent vector machine sorter after the normalization denoising.

The present invention is owing to be the distance images amplitude by an amplitude, phase place is average two spectrums that the deformation distance picture of distance images phase place twice calculates the test target distance images, and recover mean distance picture after the denoising by average two spectrum, compared with prior art, recover mean distance picture after the denoising, and then improved the object recognition rate under the low signal-to-noise ratio condition.

Description of drawings

Fig. 1 is a process flow diagram of the present invention;

Fig. 2 is mean distance picture after the denoising of two spectrum, add make an uproar mean distance picture and nothing makes an uproar mean distance as comparison diagram;

Fig. 3 is the discrimination comparison diagram that carries out Target Recognition with the power spectrum characteristic of mean distance picture before and after the denoising.

Embodiment

With reference to Fig. 1, performing step of the present invention is as follows:

Step 1, the weight coefficient of calculating linearly dependent vector machine sorter.

1a) from the radar return database, take out multi-class targets, use the method for matched filtering that its echo is carried out pulse compression, obtain the distance images sample x of training objective as training objective.

The method of matched filtering is the universal method of pulse compression.

1b) to the distance images sample x of training objective, adopt 2-norm intensity method for normalizing to eliminate its strength sensitive, obtain the training objective distance images sample z after the normalization:

In actual application environment, owing to be subjected to all multifactor influences such as target range radar distance, the distance images sample intensity of training objective differs, and this specific character is called strength sensitive; The signal to noise ratio (S/N ratio) of training objective distance images sample can reach more than the 30dB, its noise section can be ignored substantially, so for decent x of the initial distance of a training objective, adopt 2-norm intensity method for normalizing to eliminate its strength sensitive, obtain the training objective distance images sample z after the normalization:

$z=\frac{x}{\left|\right|x\left|\right|},$

Wherein, ‖ x ‖ is the 2-norm of decent x of training objective initial distance.

1c) the distance images sample z with the training objective after the normalization is expressed as { z (i), i=0,1 ..., N-1), wherein, z (i) is an i dimension element among the z, and N is a distance images sample dimension, the training objective distance images sample z after the normalization is carried out Fourier change, and get its mould value square, obtain the power spectrum characteristic of z:

D _a=(D _a(0)，D _a(1)，…，D _a(p)，…D _a(K-1)}，

Wherein, D _a(p) be training sample power spectrum characteristic D _aP dimension element, p=0,1 ..., K-1, K represent the dimension of power spectrum characteristic;

1d) use the power spectrum characteristic training linear associated vector machine sorter of training sample, obtain the weight coefficient of linearly dependent vector machine sorter: W={ ω (0), ω (1) ..., ω (q) ... ω (C-1) },

Wherein: ω (q)=ω (0, q), ω (1, q) ..., ω (K-1, q) } ^TBe the q dimension element of weight coefficient, q=0,1 ..., C-1, C are target classification number,

What use in the example of the present invention is linearly dependent vector machine sorter, but does not limit to this sorter, can also use the nonlinear dependence vector machine classifier, support vector machine classifier etc.

Step 2 is obtained the distance images sample of test target.

Radar system as test target, and is carried out pulse compression respectively to R continuous echo of this test target with detected certain unknown object, obtains test target distance images sample set X={x ₁, x ₂..., x _d..., x _R, d=1 wherein, 2 ..., R, x _dEcho carries out the distance images sample that pulse compression obtains for d time.

Step 3, mean distance picture after the denoising of acquisition test target.

3a) take out distance images sample in the test target distance images sample set successively, and the distance images sample x to taking out _d={ x _d(0), x _d(1) ... x _d(e) ... x _d(N-1) }, use following formula to produce the deformation distance picture:

y _d＝{y _d(0),y _d(1)，…y _d(e)…y _d(N-1)}:

Wherein,

Be expressed as y _dIn e dimension element, e=0,1 ..., N-1, x _d(e) be x _dIn e dimension element, e=0,1 ..., N-1, N are the distance images dimension, || the mould value is asked in expression;

3b) decent the x that adjust the distance _dDo Fast Fourier Transform (FFT), obtain x _dFrequency spectrum S _d, to deformation distance as y _dDo Fast Fourier Transform (FFT), obtain y _dFrequency spectrum T _d

3c) pass through x _dFrequency spectrum S _d, y _dFrequency spectrum T _dObtain x _dTwo spectrum B _d, this pair spectrum B _dBe N dimension square formation, B _dIn p+1 capable, q+1 column element: B _d(p, q)=S _d(p) S _d(q) T _d ^*(p+q), S wherein _d(p) be S _dP dimension element, S _d(q) be S _dQ dimension element, T _d ^*(p+q) be T _dIn the conjugation of p+q dimension element, p=0,1 ..., N-1, q=0,1 ..., N-1;

3d) by step 3c) R that obtains two spectrums, constitute two spectrum signature collection B=(B ₁, B ₂... B _d..., B _R, calculate average two spectrum signature D=1 wherein, 2 ..., R, B _dBe two spectrum signatures of the d time distance images sample;

3c) by average two spectrum signature B ', obtain the amplitude U and the phase place V of mean distance picture frequency spectrum after the denoising, its concrete steps are as follows:

3c1) be averaged the amplitude G of two spectrum signature B ', and mean distance picture frequency spectral amplitude U is that element is 1 N dimensional vector entirely after the initialization denoising, initialization deformation distance picture frequency spectral amplitude F is that element is 1 N dimensional vector entirely, and initialization iterations s is 0;

3c2) upgrade each element of U successively: the dimension element U of the q among the U (q) is updated to;

$U\left(q\right)=\frac{1}{N}\{\frac{G(q,0)F\left(q\right)}{U\left(0\right)}+\frac{G(q,1)F(q+1)}{U\left(1\right)}+\·\·\·\frac{G(q,j)F(q+j)}{U\left(j\right)}+\·\·\·+\frac{G(q,N-1)F(q+N-1)}{U(N-1)}\},$ Q=0 wherein, 1; ..N-, (q is that the q+1 of G is capable j) to G, and j+1 column element, F (q+j) are the q+j dimension element of F, and U (j) be that the j of U ties up element;

3c3) upgrade each element of F successively: the dimension element F of the q among the F (q) is updated to:

$F\left(q\right)=\frac{1}{N}\{\frac{U\left(0\right)U\left(q\right)}{G(0,q)}+\frac{U\left(1\right)U(q-1)}{G(1,q-1)}+\·\·\·+\frac{U\left(j\right)U(q-j)}{G(j,q-j)}+\·\·\·+\frac{U(N-1)U(q-N+1)}{G(N-1,q-N+1)}\},$

Q=0 wherein, 1 ... N-1, G (j, q-j) be G j+1 capable, q-j+1 column element, U (j) are the j dimension element of U; U (q-j) is for removing the q-j dimension element of U;

3c4) iterations is updated to s=s+1, if step 3c2 is then returned in s＜100 after upgrading), otherwise, enter step 3c5)

3c5) be averaged the phase place H of two spectrum signature B ', and mean distance picture frequency spectrum phase place V is that element is 1 N dimensional vector entirely after the initialization denoising, initialization deformation distance picture frequency spectrum phase place T is that element is 1 N dimensional vector entirely, and initialization iterations s is 0;

3c6) upgrade each element of V successively, the q dimension element V (q) that is about among the V is updated to:

$V\left(q\right)=\frac{1}{N}\{\frac{H(q,0)}{V\left(0\right)T\left(q\right)}+\frac{H(q,1)}{V\left(1\right)T\left(q\right)}+\·\·\·+\frac{H(q,j)}{V\left(j\right)T(q+1)}+\·\·\·+\frac{H(q,N-1)}{V(N-1)T(q+N-1)}\}$

Q=0 wherein, 1 ... N-1, (q is that the q+1 of H is capable j) to H, and j+1 column element, V (j) they are the j dimension element of V, and T (q+j) be that the q+j of T ties up element;

3c7) upgrade each element of T successively, the q dimension element T (q) that is about among the T is updated to

$T\left(q\right)=\frac{1}{N}\{\frac{H(0,q)}{V\left(0\right)T\left(q\right)}+\frac{H(1,q-1)}{V\left(1\right)V(q-1)}+\·\·\·+\frac{H(j,q-j)}{V\left(j\right)V(q-j)}+\·\·\·+\frac{H(N-1,q-N+1)}{V(N-1)V(q-N+1)}\}$

Q=0 wherein, 1 ... N-1, (j is that the j+1 of H is capable q-j) to H, and q-j+1 column element, V (j) they are the j dimension element of V, and V (q-j) be that the q-j of V ties up element;

3c8) iterations is updated to s=s+1,, returns step 3c6 if upgrade back s＜100), otherwise, enter step 3d);

That use in the example of the present invention is 3c) described in method obtain the amplitude U and the phase place V of mean distance picture frequency spectrum after the denoising, but be not limited to this method, can also use the additive method that solves the incomplete data problem.

3d) by the amplitude U and the phase place V of mean distance picture frequency spectrum after the denoising, obtain after the denoising mean distance as x '=| IFFT (UV) |, wherein | IFFT () | represent affected inverse fast fourier transform and delivery value.

Step 4 obtains the target category label.

4a) adopt 2-norm intensity method for normalizing normalization denoising after mean distance as x ', obtain mean distance picture after the normalized denoising

As x ', still have the problem of strength sensitive with mean distance after eliminating the denoising that is obtained by step 3, ‖ x ' ‖ is the 2-norm of training sample x ' in the formula;

4b) with mean distance after the denoising after the normalization as z ' be expressed as z ' (i), i=0,1 ..., N-1}, wherein, z ' is an i dimension element among the z ' (i), N is a distance images sample dimension; Normalized distance images sample z ' is carried out Fourier changes, and get its mould value square, obtain the power spectrum characteristic of z ':

D _b={ D _b(0), D _b(I) ..., D _b(p) ... D _b(K-1) }, wherein, D _b(p) be power spectrum characteristic D _bP dimension element, p=0,1 ..., K-1, K represent the dimension of power spectrum characteristic;

4c) with the power spectrum characteristic D of z ' _bBe input in the linearly dependent vector machine sorter that trains, calculate the output of sorter: y=D by its weight coefficient W _bW is because power spectrum characteristic D _bBe the K dimensional vector, weight coefficient W is the matrix of KC dimension, so the output y of sorter is the C dimensional vector;

4d) with the output y={y (0) of sorter, y (1) ..., y (g) ... y (C-1) in maximal value element corresponding class label as the target category label.For example, if y (0) be maximal value element among the y, then the target classification is 1 class, is maximal value element among the y as if y (g), and then the target classification is the g+1 class, g=0 wherein, and 1 ..., C-1.

Effect of the present invention can further specify in conjunction with emulation experiment.

1. simulated conditions

Emulation experiment is to carry out in MATLAB7.0 software, and used data are the real data that radar collects, and comprises three class aircraft: Ya Ke-42 aircrafts, diploma aircraft and amp-26 aircrafts, and noise is very little in these data, can ignore.

2. experiment content

Experiment one: be the effect of check distance images denoising, select amp-26 High Range Resolution sample of 10 Continuous Observation, these 10 distance images sample standard deviations are added the noise of 0dB, obtain adding the distance images after making an uproar.Use method of the present invention to obtain mean distance picture after the denoising, and mean distance picture after the denoising is not looked like to compare with having the mean distance picture of making an uproar and adding the mean distance of making an uproar, the result as shown in Figure 2.Wherein

Fig. 2 (a) expression does not have the mean distance picture of making an uproar;

Mean distance picture after Fig. 2 (b) expression denoising;

Fig. 2 (c) expression adds the mean distance picture of making an uproar.

Experiment two: adjust the distance after the picture denoising to the improvement of recognition performance in order to verify the present invention, select 50 distance images samples of Ya Ke-42 aircraft, diploma aircraft and amp-26 aircrafts respectively, totally 150 samples are as the distance images sample of training objective; Select 20000 distance images sample of 12000 distance images samples of Ya Ke-42 aircraft, 20000 distance images samples of diploma aircraft, amp-26 aircrafts, totally 52000 samples, in these samples, add 0dB respectively, 5dB, 10dB, the noise of 15dB and 20dB is as the distance images sample of test target.

Use the inventive method to carry out denoising in per 10 the distance images samples in the distance images sample of 52000 test targets, obtain after 1 denoising decent of mean distance, finally obtain after 5200 denoisings decent of mean distance, use the power spectrum characteristic of decent of mean distance after these 5200 denoisings to carry out experiment for target identification; Again that per 10 the distance images samples in the distance images sample of 52000 test targets are directly average, obtain 1 and add decent of the mean distance of making an uproar, finally obtain 5200 and add decent of the mean distance of making an uproar, use these 5200 power spectrum characteristics that add decent of the mean distance of making an uproar to carry out experiment for target identification.The result as shown in Figure 3.

Among Fig. 3, solid line represents to use the power spectrum characteristic that adds the mean distance picture of making an uproar to carry out the variation of the discrimination of Target Recognition with signal to noise ratio (S/N ratio), and dotted line represents to use that the power spectrum characteristic of mean distance picture carries out the variation of the discrimination of Target Recognition with signal to noise ratio (S/N ratio) after the denoising.

2. interpretation:

As seen from Figure 2: 1) do not have the mean distance picture of making an uproar, the amplitude of noise range is zero substantially, and the distance images of noise range is smoother; 2) add the mean distance picture of making an uproar, the amplitude of signaling zone slightly changes, and the amplitude integral body of noise range has raising, and the distance images of noise range is unsmooth; 3) the mean distance picture of two spectrum denoisings, the amplitude of signaling zone slightly changes, and largest peaks is slightly higher than there not being the distance images of making an uproar, it is lower slightly that less peakedness ratio does not have the distance images of making an uproar, the amplitude integral body of noise range slightly improves, but improves not obviously, and the distance images of noise range is unsmooth.

As seen from Figure 3: when signal to noise ratio (S/N ratio) is 0dB, use the power spectrum characteristic that adds the mean distance picture of making an uproar to discern fully, and the power spectrum characteristic of mean distance picture has improved 30 percentage points after the use denoising; When signal to noise ratio (S/N ratio) was in 5dB～15dB scope, the discrimination of the power spectrum characteristic of mean distance picture exceeded 10～15 percentage points than the discrimination that uses the power spectrum characteristic that adds the mean distance picture of making an uproar after the use denoising; Use during for 20dB the discrimination of the power spectrum characteristic that adds the mean distance picture of making an uproar nearly to return to 80% in signal to noise ratio (S/N ratio), and the discrimination that uses the power spectrum characteristic of mean distance picture after the denoising is than 80% height slightly.

To sum up, the present invention can realize the squelch of distance images under the low signal-to-noise ratio condition, has recognition effect preferably.

Claims (6)

1. one kind based on radar target identification method under the noise background of two spectrum denoisings, comprises the steps:

(1) from the radar return database, takes out multi-class targets as training objective, and its echo carried out pulse compression, obtain the distance images sample x of training objective, and decent the x normalization of adjusting the distance, by the weight coefficient W of distance images sample calculation linearly dependent vector machine sorter after the normalization;

(2) radar system as test target, and is carried out pulse compression respectively to R continuous echo of this test target with detected certain unknown object, obtains test target distance images sample set: X={x ₁, x ₂..., x _d..., x _R), d=1 wherein, 2 ..., R, x _dEcho carries out the distance images sample that pulse compression obtains for d time;

(3) mean distance picture after the acquisition denoising:

3a) take out distance images sample in the test target distance images sample set successively, obtain its pair spectrum:

3a1) distance images sample x to taking out _d={ x _d(0), x _d(1) ... x _d(e) ... x _d(N-1) }, use following formula to produce deformation distance picture: y _d={ y _d(0), y _d(1) ... y _d(e) ... y _d(N-1) }:

$y_d\left(e\right)=\frac{x_d\left(e\right)\·x_d\left(e\right)}{\left|x_d\left(e\right)\right|},$

X wherein _d(e) be x _dIn e dimension element, e=O, 1 ..., N-1, N are the distance images dimension;

y _d(e) be y _dIn e dimension element, e=0,1 ..., N-1, || the mould value is asked in expression;

3a2) decent the x that adjust the distance _dDo Fast Fourier Transform (FFT), obtain x _dFrequency spectrum S _d, to deformation distance as y _dDo Fast Fourier Transform (FFT), obtain y _dFrequency spectrum T _d

3a3) pass through x _dFrequency spectrum S _d, y _dFrequency spectrum T _dObtain x _dTwo spectrum B _d, this pair spectrum B _dBe N dimension square formation, B _dIn p+1 capable, q+1 column element: B _d(p, q)=S _d(p) S _d(q) T _d ^*(p+q), S wherein _d(p) be S _dP dimension element, S _d(q) be S _dQ dimension element, T _d ^*(p+q) be T _dIn the conjugation of p+q dimension element, p=0,1 ..., N-1, q=0,1 ..., N-1;

3b) by step 3a) R that obtains two spectrums, constitute two spectrum signature collection B=(B ₁, B ₂... B _d..., B _R, calculate average two spectrum signature

D=1 wherein, 2 ..., R, B _dBe two spectrum signatures of the d time distance images sample;

3c) by average two spectrum signature B ', obtain mean distance picture frequency spectral amplitude U and phase place V after the denoising;

3d) amplitude U and the phase place V to mean distance picture frequency spectrum after the denoising carries out inverse fast fourier transform, obtain denoising after mean distance as x ';

(4) mean distance obtains the target category label as x ' by the weight coefficient W of mean distance picture after the normalized denoising and linearly dependent vector machine sorter after the normalization denoising.

2. radar target identification method under the noise background according to claim 1,2-norm intensity normalization method is adopted in wherein described decent the x normalization of adjusting the distance of step 1), obtains the training sample after the normalization

Wherein || x|| is the 2-norm of training distance images sample x.

3. radar target identification method under the noise background according to claim 1, the described weight coefficient W of step 1) wherein by distance images sample calculation linearly dependent vector machine sorter after the normalization, as follows:

(1a) the training sample z after the normalization is carried out Fourier and changes, and get its mould value square, obtain the power spectrum characteristic of z: D _a=(D _a(0), D _a(1) ..., D _a(p) ... D _a(K-1) },

D wherein _a(p) be training sample power spectrum characteristic D _aP dimension element, p=0,1 ..., K-1, K are the dimension of power spectrum characteristic;

(1b) the power spectrum characteristic D of training sample after the use normalization _aTraining linear associated vector machine sorter obtains the weight coefficient of linearly dependent vector machine sorter: W={ ω (0), ω (1) ..., ω (q) ... ω (C-1) },

Wherein: ω (q)=ω (0, q), ω (1, q) ..., ω (K-1, q) } ^TBe the q dimension element of weight coefficient,

Q=0,1 ..., C-1, C are training objective classification number.

4. radar target identification method under the noise background according to claim 1, wherein step 3c) described in obtain mean distance picture frequency spectral amplitude U after the denoising, carry out as follows:

3c1) be averaged the amplitude G of two spectrum signature B ', and mean distance picture frequency spectral amplitude U is that element is 1 N dimensional vector entirely after the initialization denoising, deformation distance picture frequency spectral amplitude F is that element is 1 N dimensional vector entirely after the initialization denoising, and initialization iterations s is 0;

3c2) upgrade each element of U successively: the dimension element U of the q among the U (q) is updated to

$U\left(q\right)=\frac{1}{N}\{\frac{G(q,0)F\left(q\right)}{U\left(0\right)}+\frac{G(q,1)F(q+1)}{U\left(1\right)}+\·\·\·\frac{G(q,j)F(q+j)}{U\left(j\right)}+\·\·\·+\frac{G(q,N-1)F(q+N-1)}{U(N-1)}\},$ Q=O wherein, I ... N-, (q is that the q+1 of G is capable j) to G, and j+1 column element, F (q+j) they are the q+j dimension element of F, and U (j) be that the j of U ties up element;

3c3) upgrade each element of F successively: the dimension element F of the q among the F (q) is updated to

$F\left(q\right)=\frac{1}{N}\{\frac{U\left(0\right)U\left(q\right)}{G(0,q)}+\frac{U\left(1\right)U(q-1)}{G(1,q-1)}+\·\·\·+\frac{U\left(j\right)U(q-j)}{U(j,q-j)}\·\·\·+\frac{U(N-1)U(q-N+1)}{G(N-1,q-N+1)}\},$

Q=0 wherein, 1 ... N-1, G (j, q-j) be G j+1 capable, q-j+1 column element, U (j) are the j dimension element of U; U (q-j) is for removing the q-j dimension element of U;

3c4) iterations is updated to s=s+1, if step 3c2 is returned in s＜100 after upgrading), otherwise, enter step 3c5).

5. radar target identification method under the noise background according to claim 1, wherein step 3c) described in obtain mean distance picture frequency spectrum phase place V after the denoising, carry out as follows:

3c5) be averaged the phase place H of two spectrum signature B ', and mean distance picture frequency spectrum phase place V is that element is 1 N dimensional vector entirely after the initialization denoising, deformation distance picture frequency spectrum phase place T is that element is 1 N dimensional vector entirely after the initialization denoising, and initialization iterations s is 0;

3c6) upgrade each element of V successively: the dimension element V of the q among the V (q) is updated to

$V\left(q\right)=\frac{1}{N}\{\frac{H(q,0)}{V\left(0\right)T\left(q\right)}+\frac{H(q,1)}{V\left(1\right)T\left(q\right)}+\·\·\·+\frac{H(q,j)}{V\left(j\right)T(q+j)}+\·\·\·+\frac{H(q,N-1)}{V(N-1)T(q+N-1)}\}$

Q=0 wherein, 1 ... N-1, (q is that the q+1 of H is capable j) to H, and j+1 column element, V (j) they are the j dimension element of V, and T (q+j) be that the q+j of T ties up element;

3c7) upgrade each element of T successively: the dimension element T of the q among the T (q) is updated to

$T\left(q\right)=\frac{1}{N}\{\frac{H(0,q)}{V\left(0\right)V\left(q\right)}+\frac{H(1,q-1)}{V\left(1\right)V(q-1)}+\·\·\·+\frac{H(j,q-j)}{V\left(j\right)V(q-j)}+\·\·\·+\frac{H(N-1,q-N+1)}{V(N-1)V(q-N+1)}\}$

Q=0 wherein, 1 ... N-1, (j is that the j+1 of H is capable q-j) to H, and q-j+1 column element, V (j) they are the j dimension element of V, and V (q-j) be that the q-j of V ties up element;

3c8) iterations is updated to s=s+1,, returns step 3c6 if upgrade back s＜100), otherwise, enter step 3d).

6. radar target identification method under the noise background according to claim 1, wherein the described weight coefficient W by mean distance picture after the normalized denoising and linearly dependent vector machine sorter of step 4) obtains the target category label, as follows:

4a) mean distance after the normalized denoising is looked like to carry out Fourier and changes, and get its mould value square, obtain the power spectrum characteristic of mean distance picture after the normalized denoising:

D _b=(D _b(o), D _b(I) ..., D _b(p) ... D _b(K-1)), D wherein _b(p) be power spectrum characteristic D _bP dimension element, p=0,1 ..., K-1, K represent the dimension of power spectrum characteristic;

4b) with the power spectrum characteristic D of mean distance picture after the normalized denoising _bBe input in the linearly dependent vector machine sorter that trains, calculate the output of sorter: y=D by its weight coefficient W _bW is because power spectrum characteristic D _bBe the K dimensional vector, weight coefficient W is the matrix of KC dimension, so the output y of sorter is the C dimensional vector, C is a training objective classification number;

4c) determine the target category label, be about among the y maximal value element corresponding class label as the target category label according to the output of sorter.

Images