CN103065162A - SAR (Synthetic Aperture Radar) target azimuth angle estimation method based on sparse description - Google Patents

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) target azimuth angle estimation method based on sparse description, which mainly solves the problem of 180-degree indistinction caused when azimuth angle estimation is carried out in the prior art. The SAR target azimuth angle estimation method based on the sparse description is realized through the following steps of: (1) inputting training samples and test samples, intercepting a subimage which contains a target from the center of each of the training samples and the test samples, and carrying out histogram equalization on the intercepted subimage; (2) as for the training samples and the test samples, which are subjected to the histogram equalization, calculating the sparse description vectors of the test samples on a dictionary formed by all the training samples; (3) calculating the reconstruction error of the training sample which corresponds to each nonzero coefficient of the obtained sparse description vectors; and (4) selecting the azimuth angle of the training sample which corresponds to the nonzero coefficient with a minimum reconstruction error as an angle estimation for output. Compared with the prior art, the SAR target azimuth angle estimation method based on the sparse description, which is disclosed by the invention, has the advantages that 180-degree indistinction problem does not exist, the angle estimation accuracy is high and SAR target azimuth angle estimation and further SAR target recognition can be conducted.

Description

SAR azimuth of target method of estimation based on sparse description

Technical field:

The invention belongs to technical field of image processing, relate to the estimation of SAR azimuth of target, can be used for the target identification of SAR.

Background technology:

Synthetic-aperture radar SAR has been widely used in the civilian and military field because it has round-the-clock, round-the-clock ability to work, and the automatic target identification of SAR image has been subject to showing great attention to of people.Azimuth of target estimates it is the important step of carrying out the identification of SAR Image Automatic Target.The template method of identification is that the image template under each position angle mates identification with the image of target to be identified and known target.If before identification, can from the SAR image of target to be identified, estimate azimuth of target, then can effectively reduce quantity and the search time of template matches, improve precision and the efficient of target identification.

The position angle method of estimation of existing SAR target mainly can be divided three classes: the first kind: the position angle method of estimation of the minimum boundary rectangle of based target, such as document 1:Adaptive boosting for SAR automatic target recognition, IEEE Trans.Aerosp.Electron.Syst., vol.43, no.1,2007; Equations of The Second Kind: the position angle method of estimation of based target main shaft, such as document 2:Target aspect estimation from single and multi-pass SAR images.Acoustics, Speech and Signal Processing.Proceedings of the 1998 IEEE International Conference on; The 3rd class: the position angle method of estimation of the long primary edge of based target, cut apart and position angle estimation, National University of Defense technology's journal, the 5th phase of the 23rd volume, calendar year 2001 such as document 3:SAR image object.

Said method has all only utilized the geometry feature of target when carrying out the position angle estimation, do not utilize the image information of target, and it is low to estimate angular accuracy, estimates to have 180 ° of fuzzy problems among the result of angle, is unfavorable for the requirement of SAR target identification real-time.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, proposed a kind of SAR azimuth of target method of estimation based on sparse description, 180 ° of fuzzy problems that exist when avoiding the position angle to estimate.

The technical thought that realizes the object of the invention is, try to achieve the sparse description vectors of test sample book on the dictionary that is formed by all training samples by sparse description algorithm, then the reconstructed error of the corresponding training sample of each nonzero coefficient in the compute sparse description vectors selects the position angle of training sample corresponding to the nonzero coefficient of reconstructed error minimum as estimating angle result output.Its specific implementation step comprises as follows:

(1) input training sample and test sample book comprises the subimage of target from every width of cloth picture centre intercepting of training sample and test sample book, to reduce the impact that large-area ground unrest produces the position angle estimated performance in the SAR image;

(2) subimage of all interceptings carried out the standard histogram equalization, the impact that the position angle estimated performance is produced to weaken the non-uniform scattering that exists in the SAR image;

(3) with the subimage of the training sample behind histogram equalization structure dictionary A, utilize the following majorized function of orthogonal matching pursuit OMP Algorithm for Solving, obtain the sparse description vectors α of subimage y on dictionary A of the test sample book behind the histogram equalization;

argmin||α|| ₀?s.t.Aα=y

Wherein, || || ₀0 norm is got in expression, and min () is for getting minimum function;

The reconstructed error e of the training sample that (4) each nonzero coefficient is corresponding among the compute sparse description vectors α _k:

e _k=||y-Af _k(α)|| ₂，k=1,...,N ₀

Wherein, f _k(α) for other elements except k nonzero element among the sparse description vectors α are set to 0 function entirely, N ₀Be the number of nonzero element among the sparse description vectors α, i.e. degree of rarefication;

The position angle of the training sample that (5) nonzero coefficient of selection reconstructed error minimum is corresponding is as estimating as a result out output of angle:

$\mathrm{out}=H_{\mathrm{ang}}\left\{H_{\mathrm{index}}\right[\underset{k}{\arg \min }\left(e_k\right)\left]\right\}$

Wherein, H _Index(k) for obtaining the function of k nonzero element location label in α, H _Ang() is for obtaining the azimuthal function of sample.

The present invention compared with prior art has the following advantages:

1) the present invention utilizes all training samples to make up the dictionary matrix, adopts the method for sparse description to carry out the position angle estimation, estimates the angle result and does not have 180 ° of fuzzy problems.

2) the present invention takes full advantage of the image information of SAR target, and precision and the efficient of target identification, can be improved at template matches quantity and search time in the time of effectively reducing succeeding target identification in the high-precision position angle that estimates target of energy.

Description of drawings

Fig. 1 is realization flow figure of the present invention;

Fig. 2 carries out the sparse description vectors synoptic diagram that the position angle estimation obtains with the present invention to a sample;

Fig. 3 carries out the reconstructed error synoptic diagram that the position angle estimation obtains with the present invention to a sample;

Fig. 4 carries out the error result synoptic diagram that the position angle estimation obtains with the present invention to sample set.

Embodiment

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

Step 1: input training sample and test sample book, to comprise the size of target be 48 * 48 subimage to the intercepting of the center of every width of cloth image from training sample and test sample book, to reduce the impact that large-area ground unrest produces the position angle estimated performance in the SAR image.

Step 2: the subimage to intercepting carries out the standard histogram equalization, the variation range of image pixel value is adjusted to [0,1], so that all images have identical dynamic range, the impact that the position angle estimated performance is produced to weaken the non-uniform scattering that exists in the SAR image.

Step 3: with the subimage of the training sample behind histogram equalization structure dictionary A, utilize the following majorized function of orthogonal matching pursuit OMP Algorithm for Solving to obtain the sparse description vectors α of subimage y on dictionary A of the test sample book behind the histogram equalization.

argmin||α|| ₀s.t.Aα=y

Its concrete solution procedure reference:

(1)Tropp?J，Gilbert?A.Signal?recovery?from?random?measurements?via?orthogonal?matching?pursuit.IEEE?Transactions?on?Information?Theory，2007，53(12):4655-4666.

(2)Tropp?J.Greed?is?good:algorithmic?results?for?sparse?approximation.IEEE?Transactions?on?Information?Theory，2004，50(10):2231-2242.

Step 4: the reconstructed error e of the training sample that each nonzero coefficient is corresponding among the sparse description vectors α that calculates _k

e _k=||y-Af _k(α)|| ₂，k=1,...,K ₀，

Wherein, f _k(α) for other elements except k nonzero element among the sparse description vectors α are set to 0 function entirely, K ₀Be the number of nonzero element among the α, namely degree of rarefication is got K among the present invention ₀=7.

Step 5: select reconstructed error e _kThe position angle of the training sample that minimum nonzero coefficient is corresponding is as estimating as a result out output of angle.

$\mathrm{out}=H_{\mathrm{ang}}\left\{H_{\mathrm{index}}\right[\underset{k}{\arg \min }\left(e_k\right)\left]\right\},$

Wherein, H _Index(k) for obtaining the function of k nonzero element location label in α, H _Ang() is for obtaining the azimuthal function of sample.

Effect of the present invention can be verified by following emulation experiment:

1. experimental data: the MSTAR data that experiment adopts U.S. DARPA/AFRL working group to provide, select wherein the target of 7 kinds of different models of 3 classes to test, wherein the BMP2 infantry fighting vehicles have sn-9563, sn-9566, sn-c213 kind model; BTR70 panzer model is sn-C71; The T72 main battle tank has sn-132, sn-812, sn-s73 kind model.Adopt the SAR image of SAR when the angle of pitch is 17 ° of working group's recommendation as training sample, SAR image when the angle of pitch is 15 ° is as test sample book, the size of all target images is 128 * 128, and resolution is 0.3m * 0.3m, and the orientation coverage is 0 ~ 360 °.Kind and the number of samples of training sample and test sample book are as shown in table 1.

The kind of table 1 training sample and test sample book and number of samples

Software platform is MATLAB r2011a.

2. experimental result

Experiment flow is processed training sample and test sample book with the present invention according to flow process as shown in Figure 1, and provides and estimate the angle result.

Emulation 1 is that 254.4920 ° BMP2-9563 sample decomposes at the dictionary that all training samples form to the position angle with the inventive method, obtains sparse description vectors α, and the result as shown in Figure 2.

Emulation 2, the reconstructed error of the training sample that each nonzero coefficient is corresponding among the compute sparse description vectors α, the result is as shown in Figure 3.As seen from Figure 3, the reconstructed error minimum be the 2nd sample corresponding to nonzero coefficient, the position angle of this sample is 254.4920 °, matches with simulated conditions.

Emulation 3, with the present invention all samples being carried out the position angle estimates, the error result that obtains as shown in Figure 4, wherein Fig. 4 (a) is for carrying out the error result that the position angle is estimated to the BMP2 sample set, Fig. 4 (b) is for carrying out the error result that the position angle is estimated to the T72 sample set, Fig. 4 (c) is for carrying out the error result that the position angle is estimated to the BTR70 sample set, Fig. 4 (d) is for carrying out the error result that the position angle is estimated to whole test sample book collection.As seen from Figure 4, the position angle evaluated error of most samples is all in 5 °.The absolute error statistics of carrying out the position angle estimation with the present invention is as shown in table 2.

The absolute error statistics that table 2 position angle is estimated

As can be seen from Table 2, for 1365 width of cloth SAR images that are used for test, the absolute error of estimating the angle is less than 5 ° 1334 width of cloth that have, and proportion is 97.73%.Estimate the angle absolute error less than 10 ° 1357 width of cloth images that have, proportion is 99.41%.By above experimental result as seen, the present invention can the azimuthal estimation of high-precision realize target.

Claims (3)

1. SAR azimuth of target method of estimation based on sparse description may further comprise the steps:

(1) input training sample and test sample book comprises the subimage of target from every width of cloth picture centre intercepting of training sample and test sample book, to reduce the impact that large-area ground unrest produces the position angle estimated performance in the SAR image;

(2) subimage of all interceptings carried out the standard histogram equalization, the impact that the position angle estimated performance is produced to weaken the non-uniform scattering that exists in the SAR image;

(3) with the subimage of the training sample behind histogram equalization structure dictionary A, utilize the following majorized function of orthogonal matching pursuit OMP Algorithm for Solving, obtain the sparse description vectors α of subimage y on dictionary A of the test sample book behind the histogram equalization;

argmin||α|| ₀s.t.Aα=y

Wherein, || || ₀0 norm is got in expression, and min () is for getting minimum function;

The reconstructed error e of the training sample that (4) each nonzero coefficient is corresponding among the compute sparse description vectors α _k:

e _k=||y-Af _k(α)|| ₂，k=1,...,N ₀

Wherein, f _k(α) for other elements except k nonzero element among the sparse description vectors α are set to 0 function entirely, N ₀Be the number of nonzero element among the sparse description vectors α, i.e. degree of rarefication;

The position angle of the training sample that (5) nonzero coefficient of selection reconstructed error minimum is corresponding is as estimating as a result out output of angle:

$\mathrm{out}=H_{\mathrm{ang}}\left\{H_{\mathrm{index}}\right[\underset{k}{\arg \min }\left(e_k\right)\left]\right\}$

Wherein, H _Index(k) for obtaining the function of k nonzero element location label in α, H _Ang() is for obtaining the azimuthal function of sample.

2. the SAR azimuth of target method of estimation based on sparse description according to claim 1, wherein the subimage size of the described intercepting of step (1) is 48 * 48 pixels.

3. the SAR azimuth of target method of estimation based on sparse description according to claim 1, wherein the described standard histogram equalization operation of step (2) adjusts to [0 with the variation range of image pixel value, 1], so that all images has identical dynamic range.

Images