CN104899549A - SAR target recognition method based on range profile time-frequency image identification dictionary learning - Google Patents

Abstract

The invention provides an SAR target recognition method based on range profile time-frequency image identification dictionary learning. An SAR range profile time-frequency image is used as a recognition feature, so that influence of defocusing caused by target movement or low image quality caused by a factor such as a signal noise ratio on a target recognition effect is avoided; and identification dictionary learning is combined with dictionary learning and classifier training, so that feature information of radar target range profile time-frequency data can be effectively extracted. The invention is conducive to reducing a number of atoms in a dictionary and decreasing operation complexity in a sparse encoding process, and is also conducive to improving precision of sparse encoding, so as to improve recognition accuracy of a radar target. Further, azimuth angle estimation does not need to be performed on an SAR image target in a whole recognition process. Therefore, a recognition complex degree is reduced, and dependency of recognition accuracy on target azimuth angle estimation is also reduced. In addition, the invention has excellent recognition performance, and facilitates improving robust performance of radar target recognition.

Description

Based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook

Technical field

The present invention relates to Technology of Radar Target Identification field, particularly relate to a kind of SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook.

Background technology

Synthetic-aperture radar (Synthetic Aperture Radar is called for short SAR) technology, is adopt the movable radar be mounted on satellite or aircraft, obtains a kind of pulsed radar technology of the geographical band radar target image of high precision.Due to the complicated scattering mechanism in the Active Imaging feature of SAR and imaging process, the target property in SAR image and optical imagery difference are very large, and this brings a lot of difficulty to target's feature-extraction and identification.

Scientific research personnel have studied many Target Recognition Algorithms based on two-dimensional SAR image.Wherein, a kind of the most direct method be exactly direct using SAR image as feature, carry out the identification of target.Another kind of radar target identification method is based on wavelet transformation or multiscale analysis.In addition, the characteristics of image such as such as target area descriptor, shade is also for carrying out target identification.Adopt the feature of the physically based deformation of scattering center model can provide a kind of meticulous, the goal description that physics is relevant, but it needs to rely on estimating target position angle, therefore the accuracy rate of its radar target recognition is also subject to the accuracy restriction of target azimuth angular estimation, is therefore difficult to reach very good recognition performance.

But, the identification of SAR target can also adopt the method for based target Range Profile, target range similarly is the data image of one dimension, that the SAR complex pattern of target is obtained through a series of process, the concrete grammar of the Range Profile of SAR is obtained see document " Liao X J by SAR image conversion process, Runkle P, Carin L.Identification of ground targets from sequential high-range-resolution radar signatures.IEEE Trans.on Aerospace and Electronic Systems.2002, 2 (38): 1230-1242 ".Compare with the radar target identification method based on image, the advantage based on the radar target identification method of Range Profile to extract the characteristic information of target-sensor orientation dependence.In addition, when the uncooperative motion due to target or when causing target image fuzzy etc. due to factors such as signal to noise ratio (S/N ratio) are low, be difficult to prove effective based on the feature extraction of two dimensional image and identification, based on Range Profile feature extraction and identify by contrast advantageously.In the SAR image target identification of based target Range Profile, also there are some correlative studys.The One-dimensional scattering centres feature of Range Profile is adopted to carry out the identification of SAR target in some document, the shortcoming of this method is: can only extract point scattering feature, and the target complex electromagnetic features such as such as extended distance, frequency dispersion, resonance can not be extracted, thus identify limited accuracy.Also have researchist to adopt the higher-order spectrum feature of Range Profile to carry out the identification of SAR target, the shortcoming of this method is the wavefront vibration sensing of picture of adjusting the distance, and feature is stable not, and accuracy and the robustness of therefore its recognition methods are all greatly affected.

As can be seen from the research work of forefathers, how to extract the effective Range Profile feature of robust, and how to carry out dimension-reduction treatment for the characteristic extracted, and the characteristic information with radar target distinguishing ability can be retained after ensureing dimensionality reduction simultaneously, be the key of SAR target identification technology.

Summary of the invention

For the above-mentioned problems in the prior art, estimating target position angle is all needed in order to solve SAR image target identification in prior art, identify the problem of limited accuracy, the invention provides a kind of SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, it is by extracting Radar Target Using Range Profiles feature as SAR target distinguishing feature, and adopt differentiate dictionary learning combine carry out dictionary learning and sorter training, the sparse coding of test data under the dictionary obtained based on this Algorithm Learning is made to have identifiability, thus improve the accuracy of SAR target identification, and do not need to carry out target azimuth angular estimation to SAR image, can also avoid defocusing or the factor such as signal to noise ratio (S/N ratio) on the impact of target recognition effect.

For achieving the above object, present invention employs following technological means:

Based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, comprise the steps:

1) SAR image is converted to the Range Profile of SAR;

2) adopt adaptive Gaussian representation method, utilize Gaussian bases to carry out Breaking Recurrently expression to the Range Profile of SAR, and then calculate the time-frequency matrix of Range Profile of SAR;

3) for multiple different classes of known radar target, gather the SAR image of multiple known radar target respectively as training sample, and process the time-frequency matrix obtaining each training sample in each classification respectively according to step 1 ~ 2, thus arrange composing training sample time-frequency data matrix vector Y by the time-frequency matrix of each training sample of each classification;

4) based on training sample time-frequency data matrix vector Y and the classification information of each training sample, the learning parameter needed for objective function of LC-KSVD dictionary learning algorithm is tried to achieve respectively; Described learning parameter comprises original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀, initialization sorter matrix W ₀, label matrix H corresponding to training sample time-frequency data matrix vector Y, in order to represent training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between classification corresponding relation discriminating matrix Q and utilize original chemical handwriting practicing allusion quotation matrix D ₀training sample time-frequency data matrix vector Y is carried out to the sparse coefficient vector X of Its Sparse Decomposition gained;

5) by original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀with initialization sorter matrix W ₀respectively as the initial value of LC-KSVD dictionary learning algorithm learning dictionary matrix D, linear transformation matrix A and sorter matrix W, and by training sample time-frequency data matrix corresponding sparse coefficient vector X, the label matrix H of vector Y with differentiate that matrix Q substitutes in the objective function of LC-KSVD dictionary learning algorithm in the lump and carry out learning and training, obtain the study dictionary matrix D after LC-KSVD dictionary learning _new, linear transformation matrix A _newwith sorter matrix W _new;

6) study dictionary matrix D is utilized _new, linear transformation matrix A _newwith sorter matrix W _newcalculate benchmark dictionary matrix reference Transforming matrix with benchmark sorter matrix

$\hat{D}=\{\frac{d_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{d_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{d_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{d_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{A}=\{\frac{a_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{a_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{a_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{a_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{W}=\{\frac{w_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{w_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{w_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{w_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

Wherein, d _{new, k}represent study dictionary matrix D _newa kth dictionary atom, a _{new, k}represent linear transformation matrix A _newa kth column vector, w _{new, k}presentation class device matrix W _newa kth column vector, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection; || || ₂for l ₂norm operational symbol;

7) for radar target to be measured, gather the SAR image of radar target to be measured, obtain the time-frequency matrix z of radar target to be measured according to step 1 ~ 2 process, and utilize benchmark dictionary matrix adopt sparse coding learning algorithm that the time-frequency matrix z of radar target to be measured is carried out Its Sparse Decomposition:

$z=\hat{D}{x}_z;$

Obtain the sparse coefficient x that the time-frequency matrix z of radar target to be measured is corresponding _z;

8) benchmark sorter matrix is utilized determine the radar target classification j belonging to radar target to be measured:

$j=\arg \underset{j}{\max }(J=\hat{W}{x}_z);$

Wherein, J represents the class label vector that radar target to be measured is corresponding, and j ∈ J;

Realize the Classification and Identification to radar target to be measured thus.

In the above-mentioned SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, preferably, described step 2 is specially:

Adopt adaptive Gaussian representation method, utilize Gaussian bases to carry out Breaking Recurrently expression to the Range Profile of SAR:

$r\left(t\right)=\underset{i=1}{\overset{i_{\max }}{\mathrm{\Σ}}}{C}_i{g}_i\left(t\right)+r_{i_{\max }+1}\left(t\right),$ And ${\left|\right|r\left(t\right)\left|\right|}^2=\underset{i=0}{\overset{i_{\max }}{\mathrm{\Σ}}}{\left|C_i\right|}^2+{\left|\right|r_{i_{\max }+1}\left(t\right)\left|\right|}^2;$

Wherein, t represents the time; R (t) represents the Range Profile of the SAR of time t; g _it () represents that picture r (t) of adjusting the distance carries out the Gaussian bases of i-th Breaking Recurrently, C _irepresent that picture r (t) of adjusting the distance carries out the expansion coefficient of i-th Breaking Recurrently, i ∈ 1,2 ..., i _max, i _maxrepresent the iteration total degree carrying out Breaking Recurrently expression, and iteration total degree i _maxmake the reconstructed error of Breaking Recurrently meet ε is default reconstructed error threshold value, and 10 ^-3≤ ε≤10 ^-5; And have relational expression:

$g_i\left(t\right)={\left(\frac{1}{\mathrm{\π}{\mathrm{\α}}_i}\right)}^{\frac{1}{4}}\exp \{-\frac{{(t-t_i)}^2}{2{\mathrm{\α}}_i}\}\·\exp \left(j2\mathrm{\π}{f}_{i}t\right);$

${\left|C_i\right|}^2=\underset{t_i,f_i,{\mathrm{\α}}_i}{\max }{|\∫r_i\left(t\right)g_i^{*}\left(t\right)\mathrm{dt}|}^2,{\mathrm{\α}}_i\∈R^{+},t_i,f_i\∈R;$

T _i, f _i, α _irepresent that picture r (t) of adjusting the distance carries out the resolution parameter in the Gaussian bases of i-th Breaking Recurrently; r _it () represents that picture r (t) of adjusting the distance carries out the iteration discrepance before i-th Breaking Recurrently, and $r_i\left(t\right)=\left\{\begin{array}{cc}r\left(t\right),& i=1\\ r_{i-1}\left(t\right)-C_{i-1}{g}_{i-1}\left(t\right),& i\≥2\end{array};\right.$ represent Gaussian bases g _ithe conjugation of (t); || represent the operational symbol that takes absolute value; represent adjustment resolution parameter t _i, f _i, α _imake || ²maximum maximum operator;

Adopt Fourier Transform Algorithm to solve above-mentioned relation formula, obtain the resolution parameter t of i-th Breaking Recurrently _i, f _i, α _iwith expansion coefficient C _i, and then time-frequency matrix Θ (t, f) of Range Profile r (t) is obtained by following formula:

$\mathrm{\Θ}(t,f)=\underset{i=0}{\overset{i_{\max }}{\mathrm{\Σ}}}2{\left|C_i\right|}^2\exp \{-\frac{{(t-t_i)}^2}{2{\mathrm{\α}}_i}-{\left(2\mathrm{\π}\right)}^2{\mathrm{\α}}_i{(f-f_i)}^2\};$

Wherein, t represents the time, and f represents frequency.

In the above-mentioned SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, preferably, described step 4 is specially:

41) respectively K-SVD dictionary learning is carried out to the time-frequency matrix of each training sample in each classification, obtain the KSVD dictionary that the training sample of each classification is corresponding, and respectively corresponding class label is marked to the dictionary atom in the KSVD dictionary of each classification, thus form original chemical handwriting practicing allusion quotation matrix D by each set with the dictionary atom of class label of all categories ₀;

42) original chemical handwriting practicing allusion quotation matrix D is utilized ₀, adopt sparse coding learning algorithm to carry out Its Sparse Decomposition to training sample time-frequency data matrix vector Y:

Y＝D ₀X；

Obtain the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding;

43) the label matrix H that training sample time-frequency data matrix vector Y is corresponding is generated; Described label matrix H=[h ₁, h ₂..., h _k..., h _k], k ∈ 1,2 ..., K}, K represent total number of training sample, h _kfor the column vector of k row in label matrix H, represent a kth training sample time-frequency data matrix y in training sample time-frequency data matrix vector Y _kclass label vector, class label vector h _kcomprise M numerical value, M represents the classification sum of the training sample of collection, wherein only training sample time-frequency data matrix y _kaffiliated m numerical value corresponding to m class is 1, and its remainder values is 0, i.e. h _k=[0 ..., 1 ..., 0] ^t, T is transposition symbol;

44) according to training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between classification corresponding relation, generate and differentiate matrix Q; Described discriminating matrix Q=[q ₁, q ₂..., q _n..., q _n], n ∈ 1,2 ..., N}, q _nrepresent the column vector differentiating n row in matrix Q, and k ∈ 1,2 ..., K}, wherein value in order to represent training sample time-frequency data matrix vector Y in a kth training sample time-frequency data matrix y _kwith original chemical handwriting practicing allusion quotation matrix D ₀in the n-th dictionary atom d _nclassification corresponding relation, if a described kth training sample time-frequency data matrix y _kwith described n-th dictionary atom d _nbelong to identical category, then get otherwise get k represents total number of training sample, and N represents original chemical handwriting practicing allusion quotation matrix D ₀in the total number of dictionary atom that comprises, T is transposition symbol;

45) pass through based on quadratic loss function and l ₂the polynary ridge regression equation of norm, solves respectively and obtains initialization linear transformation matrix A ₀with initialization sorter matrix W ₀:

A ₀＝(XX ^T+λ ₂I) ^-1XQ ^T；

W ₀＝(XX ^T+λ ₁I) ^-1XH ^T；

Wherein, I is the unit matrix of K × K, λ ₁and λ ₂for the specific gravity control parameter of polynary ridge regression equation, T is transposition symbol.

In the above-mentioned SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, preferably, the objective function of described LC-KSVD dictionary learning algorithm is:

$<D,W,A,X>=\arg \underset{D,W,A,X}{\min }{\left|\right|Y-\mathrm{DX}\left|\right|}_2^2+{\mathrm{\&beta;}}_1{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+{\mathrm{\&beta;}}_2{\left|\right|H-\mathrm{WX}\left|\right|}_2^2;$

It meets $\&ForAll;k,{\left|\right|x_k\left|\right|}_0\&le;S_p;$

Wherein, x _krepresent the kth sparse coefficient in the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection; S _prepresent degree of rarefication threshold value; || || ₀for l ₀norm operational symbol, || || ₂for l ₂norm operational symbol; β ₁and β ₂represent sparse coefficient Error weight value and sorter Error weight value in objective function respectively.

In the above-mentioned SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, preferably, described degree of rarefication threshold value S _pspan be [1,20].

In the above-mentioned SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, preferably, described sparse coefficient Error weight value β ₁with sorter Error weight value β ₂span be (0,1].

Compared to prior art, the present invention has following beneficial effect:

1, when causing due to target travel defocusing or the factor such as signal to noise ratio (S/N ratio) causes picture quality not high, the target recognition effect of based target SAR image zooming-out is not good, but the present invention is based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, then do not limit by these.

2, the present invention is based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, dictionary learning and sorter training is carried out by differentiating that dictionary learning is combined, characteristic information when effectively can extract Radar Target Using Range Profiles in audio data, not only be conducive to reducing the atom number in dictionary, reduce the computational complexity in sparse coding process, also help the precision improving sparse coding simultaneously, thus improve the recognition accuracy to radar target.

3, the present invention is based in the whole identifying of the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, all do not need to carry out position angle estimation to SAR image target, because this reducing identification complexity, it also avoid and identify that accuracy is to the dependence of target azimuth angular estimation, combine based on Range Profile time-frequency figure simultaneously and differentiate that dictionary learning technology is carried out radar target recognition and also had good recognition performance in a noisy environment, the robust performance promoting radar target recognition can be helped.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the SAR target identification method that the present invention is based on the other dictionary learning of Range Profile time-frequency illustrated handbook.

Fig. 2 is the exemplary plot of the Range Profile of SAR.

Fig. 3 is the time-frequency diagram illustration of the Range Profile of SAR.

Fig. 4 is the time-frequency illustrated example in the embodiment of the present invention before and after over-segmentation, logarithm strengthen.

Fig. 5 is the graph of relation of discrimination and degree of rarefication in the embodiment of the present invention.

Fig. 6 learns by K-SVD the dictionary exemplary plot that obtains in the embodiment of the present invention.

Fig. 7 learns by LC-KSVD the dictionary exemplary plot that obtains in the embodiment of the present invention.

Fig. 8 be in the embodiment of the present invention when degree of rarefication threshold value value is 1 dictionary dimension to the influence curve comparison diagram of discrimination.

Fig. 9 be in the embodiment of the present invention when degree of rarefication threshold value value is 10 dictionary dimension to the influence curve comparison diagram of discrimination.

Embodiment

Below in conjunction with drawings and Examples, technical scheme of the present invention is further described.

The present invention proposes a kind of SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook, the method is by extracting Radar Target Using Range Profiles feature as SAR target distinguishing feature, adopt and differentiate dictionary learning in conjunction with sparse coding, dictionary learning and sorter training are combined carry out, then utilize the identification training the sorter obtained to carry out SAR target.

Rarefaction representation and dictionary learning:

Why rarefaction representation can represent by less data volume that the reason of mass data information is one of mass data simple and important characteristic: although the dimension of data itself is very high, some structural information in many applications for same class data has redundancy.That is, those are belonged to the data of same rank or adjacent subspace, subfamily, their distribution can be represented by a representative redundant samples set.This redundant samples set namely crosses complete dictionary.

The rarefaction representation of signal is exactly to select the dictionary atom of minority from crossing complete dictionary and to represent specific signal by the linear combination of these atoms.The element of dictionary is called as atom, normally unit vector.If set D={d _k, k=1 ..., N}, p represent the size of dictionary.Element d in D _ka unit vector becoming whole Hilbert space D, and N > > n.Now, for arbitrary n dimensional signal (i) ^-1can be represented by the linear combination of dictionary atom:

$y=\mathrm{Dx}={\mathrm{\&Sigma;}}_{k=1}^N{x}_k{d}_k---\left(7\right)$

Wherein, p is the expansion coefficient of signal, because dictionary was complete p, so coefficient vector x is not unique.In order to obtain the most sparse one from all coefficient vectors, just need to introduce sparsity constraints condition.Namely

meet y=Dx (8)

Wherein, || x|| ₀represent the l of vectorial p ₀it, norm, the namely number (degree of rarefication) of nonzero element in vector x, call that " norm " is because it is v in p norm _x=(D ^td) ^-1d ^tz _xtime the limit.But in fact it is not the norm of a real meaning, and it is unlike l ₁norm has all features of a norm like that.Want Exact Solution formula to be that more complicated is also very consuming time, in order to obtain the rarefaction representation coefficient of signal fast, usually loosen in reality the accurate requirement of expression, the approximate error η that the Sparse of namely trying to achieve represents is no more than threshold epsilon.Then above problem is just converted into:

$\arg \underset{x}{\min }{\left|\right|x\left|\right|}_0$ Meet y=Dx+ η and ${\left|\right|\mathrm{\&eta;}\left|\right|}_2^2<\mathrm{\&epsiv;}---\left(9\right)$

L ₀norm is a np problem, can only be obtained the suboptimal solution of sparse coefficient vector x at present by various approximate data.Ask the algorithm of suboptimal solution can be divided into two large classes: the first kind is greedy algorithm, if match tracing (MP) and orthogonal matching pursuit (Orthogonal Matching Pursuit, OMP) etc. are all the optimal base vectors being selected local by iteration.Equations of The Second Kind is based on convex lax global optimization approach, as base tracing algorithm (Basis Pursuit, BP), noise reduction algorithm (Basis Pursuit Denoising followed the trail of by base, BPDN) and least absolute value compression selector switch (The Least Absolute Shrinkage and Selection Operator, LASSO).

Specifically, based in the target identification of sparse representation theory, dictionary and the sorter designed how is selected to be the main two problems studied.

Dictionary learning is important research content and the study hotspot of rarefaction representation signal processing theory in recent years, based in the target identification of dictionary learning, how to select dictionary and the sorter designed to be the main two problems studied.Fixing dictionary and adaptive learning dictionary is mainly contained at dictionary design aspect.It is fixing that dictionary is alternative has Fourier dictionary, Wavelet dictionary, Gabor dictionary, Curvelet dictionary etc.Adaptive learning dictionary is the dictionary by obtaining the adaptive learning process of training sample, conventional adaptive learning dictionary training algorithm comprises K-SVD dictionary learning algorithm, ILSDLA dictionary learning algorithm, RL-DLA dictionary learning algorithm etc., wherein K-SVD dictionary learning algorithm application is comparatively general, this algorithm is the Michal Aharon by the Institute of Technology of Israel in 2006, the people such as Michael Elad put forward (see existing document " M.Aharon, M.Elad and A.Bruckstein.K-SVD:An Algorithm for Designing Overcomplete Dictionaries for Sparse Representation [J] .IEEE Transactions on Signal Processing, 2006, 54 (1): 4311-4322 "), it is a kind of very classical dictionary training algorithm.Self-adapting dictionary study can find one to have stronger rarefaction representation ability and the less dictionary of scale for signal specific.When the dictionary obtained when utilizing adaptive learning carries out target identification, if dictionary learning and sorter training are still carried out separately, can cause the dictionary that learns to obtain for target identification be not optimum.Now, dictionary learning and sorter training are combined and carry out, the complete dictionary of the mistake obtained making study is more suitable for object recognition task.This based on crossing the Target Recognition Algorithms of complete dictionary learning by adding discriminability condition in objective function, the dictionary that study is obtained has reconstruct for data and is easy to the feature of Classification and Identification.This dictionary learning theory is called discriminating dictionary learning.

The present invention is based in the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, have employed label unanimously differentiate dictionary learning algorithm combine carry out dictionary learning and sorter training.Dictionary learning algorithm (Label Consistent K-SVD unanimously differentiated by label, usually referred to as LC-KSVD dictionary learning algorithm) put forward (see existing document " Zhuolin Jiang first in 2011 by people such as Z.L.Jiang, Zhe Lin and Larry S.Davis.Label Consistent K-SVD:Learning a Discriminative Dictionary for Recognition [J] .IEEE Transactions on Patttern Analysisand Machine Intelligence.2013, 35 (11): 2651-2664 "), this algorithm not only make use of the classification information of training data, and classification information is joined in dictionary atom, the sparse coding of test data under the dictionary obtained based on this Algorithm Learning is made to have identifiability, these sparse codings are directly classified device and adopt, in order to carry out Classification and Identification to radar target to be measured, to obtain the high recognition result of accuracy rate.

The objective function of LC-KSVD:

In LC-KSVD dictionary learning algorithm, the objective function of dictionary learning is made up of reconstructed error item, error in classification item and sparse coding driscrimination error item, and its expression formula is:

$<D,W,A,X>=\arg \underset{D,W,A,X}{\min }{\left|\right|Y-\mathrm{DX}\left|\right|}_2^2+{\mathrm{\&beta;}}_1{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+{\mathrm{\&beta;}}_2{\left|\right|H-\mathrm{WX}\left|\right|}_2^2;$

It meets s _prepresent degree of rarefication threshold value.

Wherein, || || ₀for l ₀norm operational symbol, || || ₂for l ₂norm operational symbol; β ₁and β ₂represent sparse coefficient Error weight value and sorter Error weight value in objective function respectively; T is transposition symbol.

D is study dictionary matrix, and A is linear transformation matrix, and W is sorter matrix, is three parameters needing to carry out learning training in LC-KSVD dictionary learning algorithm.These three parameters need to carry out initialization in the initial step of study, and its respective initialization value can be defined as original chemical handwriting practicing allusion quotation matrix D respectively ₀, initialization linear transformation matrix A ₀with initialization sorter matrix W ₀.

X is for utilizing original chemical handwriting practicing allusion quotation matrix D ₀training sample time-frequency data matrix vector Y is carried out to the sparse coefficient vector of Its Sparse Decomposition gained, x _krepresent the kth sparse coefficient in the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection.Therefore, item illustrates reconstructed error training sample time-frequency data matrix vector Y being carried out to coefficient reconstruct.

Q represents training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between the discriminating matrix of classification corresponding relation, be used to the identifiability of the sparse coding describing training sample time-frequency data matrix vector Y.Differentiate matrix Q=[q ₁, q ₂..., q _n..., q _n], be K capable × matrix of N row, n ∈ 1,2 ..., N}, q _nrepresent the column vector differentiating n row in matrix Q, and k ∈ 1,2 ..., K}, K represent total number of training sample, and N represents original chemical handwriting practicing allusion quotation matrix D ₀in the total number of dictionary atom that comprises.Wherein, element value in order to represent training sample time-frequency data matrix vector Y in a kth training sample time-frequency data matrix y _kwith original chemical handwriting practicing allusion quotation matrix D ₀in the n-th dictionary atom d _nclassification corresponding relation, if a described kth training sample time-frequency data matrix y _kwith described n-th dictionary atom d _nbelong to identical category, then get otherwise get classification due to each training sample is known, and original chemical handwriting practicing allusion quotation matrix D ₀also be obtained through K-SVD dictionary learning by training sample time-frequency data matrix vector Y, therefore, original chemical handwriting practicing allusion quotation matrix D ₀in the classification of each dictionary atom be also known, thus can determine to differentiate each element in matrix Q value.For example, original chemical handwriting practicing allusion quotation matrix D is supposed ₀=[d ₁..., d ₆], training sample time-frequency data matrix vector Y=[y ₁..., y ₆] in, y ₁, y ₂, d ₁and d ₂belong to the first kind, y ₃, y ₄, d ₃and d ₄belong to Equations of The Second Kind, y ₅, y ₆, d ₅and d ₆belong to the 3rd class, then Q can be defined as:

$Q=\left(\begin{array}{cccccc}1& 1& 0& 0& 0& 0\\ 1& 1& 0& 0& 0& 0\\ 0& 0& 1& 1& 0& 0\\ 0& 0& 1& 1& 0& 0\\ 0& 0& 0& 0& 1& 1\\ 0& 0& 0& 0& 1& 1\end{array}\right).$

A is a linear transformation matrix, and it is treated as linear transformation g (x, A)=Ax herein, original sparse coding coefficient x is transformed into the sparse features space with maximum identifiability by it xiang Ze illustrates and can differentiate sparse coding error, and it makes sparse coding X be similar to can differentiate sparse coding Q.It makes uniformity signal have closely similar rarefaction representation, impels the sparse coding finally obtained to have tag compliance like this, to realize utilizing linear classifier to obtain the high recognition result of accuracy rate.

illustrate error in classification.Parameter W presentation class device.H is the label matrix that training sample time-frequency data matrix vector Y is corresponding, H=[h ₁, h ₂..., h _k..., h _k], M is capable × matrix of K row, k ∈ 1,2 ..., K}, h _kfor the column vector of k row in label matrix H, represent a kth training sample time-frequency data matrix y in training sample time-frequency data matrix vector Y _kclass label vector, class label vector h _kcomprise M numerical value, M represents the classification sum of the training sample of collection, wherein only training sample time-frequency data matrix y _kaffiliated m numerical value corresponding to m class is 1, and its remainder values is 0, i.e. h _k=[0 ..., 1 ..., 0] ^t.Classification due to each training sample is known, and therefore label matrix H also directly can determine according to the classification situation of each training sample in training sample time-frequency data matrix vector Y.

Suppose differentiate sparse coding X'=AX and be reversible, then have D'=DA ^-1, W'=WT ^-1.Formula just can be rewritten into:

$<D^{\&prime;},W^{\&prime;},X^{\&prime;}>=\arg \underset{D^{\&prime;},W^{\&prime;},X^{\&prime;}}{\min }{\left|\right|Y-D^{\&prime;}{X}^{\&prime;}\left|\right|}_2^2+\mathrm{\&alpha;}{\left|\right|Q-X^{\&prime;}\left|\right|}_2^2+\mathrm{\&beta;}{\left|\right|H-W^{\&prime;}{X}^{\&prime;}\left|\right|}_2^2;$

Meet $\&ForAll;k,{\left|\right|x_k\left|\right|}_0\&le;S_p;$

Wherein, Section 1 represents reconstructed error, and Section 2 is sparse coding identification error, and Section 3 is error in classification.Section 2 all kinds of sparse codings is made to have distinctive, Section 3 study is impelled to obtain an optimum sorter.

The inner structure that the dictionary learnt by the method is applicable to training data (has strict sparse constraint to each data in training set, be conducive to their rarefaction representation), no matter and dictionary be greatly little it can both produce there is distinctive sparse coding.These sparse codings directly can be classified device and adopt.This sparse coding can identification feature great for the influential effect of linear classifier.

The study solution procedure of LC-KSVD objective function:

In the training process, first through type calculates D, A and X, and then training classifier W.

The Optimization Solution of LC-KSVD dictionary learning algorithm can be converted into K-SVD algorithm to solve.Then the objective function of LC-KSVD dictionary learning algorithm can be rewritten into:

$<D,W,A,X>=\underset{D,W,A,X}{\arg \min }{\left|\right|\left(\begin{array}{c}Y\\ \sqrt{\mathrm{\&alpha;}}Q\\ \sqrt{\mathrm{\&beta;}}H\end{array}\right)-\left(\begin{array}{c}D\\ \sqrt{\mathrm{\&alpha;}}A\\ \sqrt{\mathrm{\&beta;}}H\end{array}\right)\left|\right|}_2^2---\left(12\right)$

Meet $\&ForAll;k,{\left|\right|x_k\left|\right|}_0\&le;S_p;$

If $Y_{\mathrm{new}}={(Y^T,\sqrt{\mathrm{\&alpha;}}{Q}^T,\sqrt{\mathrm{\&beta;}}{H}^T)}^T,D_{\mathrm{new}}={(D^T,\sqrt{\mathrm{\&alpha;}}{A}^T,\sqrt{\mathrm{\&beta;}}{W}^T)}^T.$ Matrix D _newl ₂row normalization under norm.The optimization of LC-KSVD dictionary learning algorithm objective function is equivalent to following problem:

$<D_{\mathrm{new}},X>=\arg \underset{D_{\mathrm{new}},X}{\min }\left\{{\left|\right|Y_{\mathrm{new}}-D_{\mathrm{new}}X\left|\right|}_2^2\right\}$

Meet $\&ForAll;k,{\left|\right|x_k\left|\right|}_0\&le;S_p;$

In fact this is exactly the problem that K-SVD algorithm solves.According to K-SVD algorithm, d _kand corresponding coefficient (row k in sparse coefficient vector X) upgrades simultaneously.If with with represent respectively and E _kin nonzero element.D _kwith can be calculated by following formula:

$<d_k,{\tilde{x}}_R^k>=\arg \underset{d_k,{\tilde{x}}_R^k}{\min }\left\{{\left|\right|{\tilde{E}}_k-d_k{\tilde{x}}_R^k\left|\right|}_F^2\right\}---\left(14\right)$

Right carry out SVD decomposition: just d can be obtained _kwith

$\begin{array}{cc}{d}_k=U(:,1)& {\tilde{x}}_R^k\end{array}=\mathrm{\&Sigma;}\left(\mathrm{1,1}\right)V(:,1)---\left(15\right)$

Finally, use replace in nonzero element.

LC-KSVD algorithm learns D simultaneously, A and W avoids the local extremum in Multiple Optimization problem and this algorithm is also applicable to the more situation of classification number.In addition, in the algorithm, in objective function, add one differentiate sparse coding error.This makes the discriminating sparse coding of same item signal have similar rarefaction representation, and this is very crucial for target identification.

The parameter initialization of LC-KSVD:

In LC-KSVD dictionary learning algorithm, need to carry out initialization to study dictionary matrix D, linear transformation matrix A and the sorter matrix W in the objective function of LC-KSVD dictionary learning algorithm, therefore need first to ask for obtaining original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀with initialization sorter matrix W ₀.

For D ₀, separately K-SVD dictionary learning is carried out to each class, the dictionary learning to obtain is joined together.Atom d in each class K-SVD dictionary _kinit Tag be obtained by its classification.In dictionary learning process, although dictionary atom can be updated, what the atomic tag after upgrading represented is also identical classification.That is be changeless at the class label of follow-up whole dictionary learning process Atom.D ₀in the number of each class atom according to dictionary D ₀total atom number and atom classification number pro-rata.

At initialization A ₀time, by following based on quadratic loss function and l ₂the polynary ridge regression model of norm:

$A=\arg \underset{A}{\min }{\left|\right|Q-\mathrm{AX}\left|\right|}^2+{\mathrm{\&lambda;}}_2{\left|\right|A\left|\right|}_2^2---\left(16\right)$

Obtain following solution:

A＝(XX ^T+λ ₂I) ^-1XQ ^T(17)

For W ₀, separated as follows by same ridge regression model:

W＝(XX ^T+λ ₁I) ^-1XH ^T(18)

There is initialized D ₀, the sparse coding X of calculation training data Y just can be carried out by original K-SVD algorithm.Then by calculating A with formula ₀and W ₀.

Sorting criterion based on LC-KSVD:

By the study dictionary matrix D obtained after LC-KSVD dictionary learning _new, linear transformation matrix A _newwith sorter matrix W _newcan be expressed as:

D _new＝{d _new,1,d _new,2,…,d _new,k…,d _new,K}；

A _new＝{a _new,1,a _new,2,…,a _new,k…,a _new,K}；

W _new＝{w _new,1,w _new,2,…,w _new,k…,w _new,K}；

D _{new, k}represent study dictionary matrix D _newa kth dictionary atom, a _{new, k}represent linear transformation matrix A _newa kth column vector, w _{new, k}presentation class device matrix W _newa kth column vector, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection.

But directly can not utilize D _new, A _newand W _newtest, because in LC-KSVD algorithm, three W is at D _newin be associating L ₂norm normalization, namely for any k, has required benchmark dictionary matrix reference Transforming matrix with benchmark sorter matrix can pass through to calculate as follows:

$\hat{D}=\{\frac{d_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{d_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{d_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{d_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{A}=\{\frac{a_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{a_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{a_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{a_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};---\left(19\right)$

$\hat{W}=\{\frac{w_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{w_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{w_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{w_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

For a test data z (the time-frequency matrix of radar target to be measured), first calculated its sparse coding x by following formula _z:

$x_z=\arg \underset{x_z}{\min }\left\{{\left|\right|z-\hat{D}{x}_z\left|\right|}_2^2\right\}---\left(20\right)$

Meet $\&ForAll;k,{\left|\right|x_z\left|\right|}_0\&le;S_p;$

Namely Its Sparse Decomposition is carried out to test data z obtain the sparse coefficient x that the time-frequency matrix z of radar target to be measured is corresponding _z, then, pass through linear classifier estimate the class label j of test data z:

$j=\arg \underset{j}{\max }(J=\hat{W}{x}_z)---\left(21\right)$

Wherein, J represents the class label vector that radar target to be measured is corresponding, and j ∈ J.

Based on above-mentioned mentality of designing, the overall flow that the present invention is based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook as shown in Figure 1, specifically comprises the steps:

1) SAR image is converted to the Range Profile of SAR.

Range Profile is the data image of one dimension, that the SAR complex pattern of target is obtained through a series of process, the concrete grammar being obtained the Range Profile of SAR by SAR image conversion process is prior art, at documents and materials " Liao X J, Runkle P, Carin L.Identification of ground targets from sequential high-range-resolution radar signatures.IEEE Trans.on Aerospace and Electronic Systems.2002, 2 (38): 1230-1242 " comparatively detailed introduction is had in, the present invention repeats no more.The example of the Range Profile of the SAR that conversion process obtains as shown in Figure 2.

2) adopt adaptive Gaussian representation method, utilize Gaussian bases to carry out Breaking Recurrently expression to the Range Profile of SAR, and then calculate the time-frequency matrix of Range Profile of SAR.

In prior art group, there is the time frequency analyzing tool of many signals, such as Short Time Fourier Transform, Wigner-Willie distribution (Wigner-ville Distribution, be abbreviated as WVD), wavelet analysis, adaptive Gaussian representation (Adaptive Gaussian Representation, is abbreviated as AGR) etc.But not all time-frequency analysis technology is all suitable for the feature extraction of Radar Target Using Range Profiles.The present invention selects AGR adaptive Gaussian representation method method to carry out the time frequency analysis of Range Profile.Its reason is, compared to other Time-Frequency Analysis Method, the time-frequency distributions local center that AGR obtains just can be corresponding to the scattering mechanism such as scattering center and local resonance phenomenon; Meanwhile, AGR can provide the associating non-negative time-frequency distributions of Range Profile, and is adaptive, and does not have cross term interference.

This step is specially:

Adopt adaptive Gaussian representation method, utilize Gaussian bases to carry out Breaking Recurrently expression to the Range Profile of SAR:

$r\left(t\right)=\underset{i=1}{\overset{i_{\max }}{\mathrm{\&Sigma;}}}{C}_i{g}_i\left(t\right)+r_{i_{\max }+1}\left(t\right),$ And ${\left|\right|r\left(t\right)\left|\right|}^2=\underset{i=0}{\overset{i_{\max }}{\mathrm{\&Sigma;}}}{\left|C_i\right|}^2+{\left|\right|r_{i_{\max }+1}\left(t\right)\left|\right|}^2;$

Wherein, t represents the time; R (t) represents the Range Profile of the SAR of time t; g _it () represents that picture r (t) of adjusting the distance carries out the Gaussian bases of i-th Breaking Recurrently, C _irepresent that picture r (t) of adjusting the distance carries out the expansion coefficient of i-th Breaking Recurrently, i ∈ 1,2 ..., i _max, i _maxrepresent the iteration total degree carrying out Breaking Recurrently expression, and iteration total degree i _maxmake the reconstructed error of Breaking Recurrently meet ε is default reconstructed error threshold value, and 10 ^-3≤ ε≤10 ^-5; And have relational expression:

$g_i\left(t\right)={\left(\frac{1}{\mathrm{\&pi;}{\mathrm{\&alpha;}}_i}\right)}^{\frac{1}{4}}\exp \{-\frac{{(t-t_i)}^2}{2{\mathrm{\&alpha;}}_i}\}\&CenterDot;\exp \left(j2\mathrm{\&pi;}{f}_{i}t\right);$

${\left|C_i\right|}^2=\underset{t_i,f_i,{\mathrm{\&alpha;}}_i}{\max }{|\&Integral;r_i\left(t\right)g_i^{*}\left(t\right)\mathrm{dt}|}^2,{\mathrm{\&alpha;}}_i\&Element;R^{+},t_i,f_i\&Element;R;$

T _i, f _i, α _irepresent that picture r (t) of adjusting the distance carries out the resolution parameter in the Gaussian bases of i-th Breaking Recurrently; r _it () represents that picture r (t) of adjusting the distance carries out the iteration discrepance before i-th Breaking Recurrently, and have:

$r_i\left(t\right)=\left\{\begin{array}{cc}r\left(t\right),& i=1\\ r_{i-1}\left(t\right)-C_{i-1}{g}_{i-1}\left(t\right),& i\&GreaterEqual;2\end{array};\right.$

represent Gaussian bases g _ithe conjugation of (t); || represent the operational symbol that takes absolute value; represent adjustment resolution parameter t _i, f _i, α _imake || ²maximum maximum operator;

Adopt Fourier Transform Algorithm to solve above-mentioned relation formula, obtain the resolution parameter t of i-th Breaking Recurrently _i, f _i, α _iwith expansion coefficient C _i, and then time-frequency matrix Θ (t, f) of Range Profile r (t) is obtained by following formula:

$\mathrm{\&Theta;}(t,f)=\underset{i=0}{\overset{i_{\max }}{\mathrm{\&Sigma;}}}2{\left|C_i\right|}^2\exp \{-\frac{{(t-t_i)}^2}{2{\mathrm{\&alpha;}}_i}-{\left(2\mathrm{\&pi;}\right)}^2{\mathrm{\&alpha;}}_i{(f-f_i)}^2\}.$

Wherein, t represents the time, and f represents frequency.

The time-frequency illustrated example of the Range Profile of the SAR obtained thus as shown in Figure 3.

3) for multiple different classes of known radar target, gather the SAR image of multiple known radar target respectively as training sample, and process the time-frequency matrix obtaining each training sample in each classification respectively according to step 1 ~ 2, thus arrange composing training sample time-frequency data matrix vector Y by the time-frequency matrix of each training sample of each classification.

This step mainly obtains the time-frequency matrix of multiple known class training sample, and the known radar target of each classification has all gathered multiple training sample tries to achieve its time-frequency matrix, in order to build study dictionary.

4) based on training sample time-frequency data matrix vector Y and the classification information of each training sample, the learning parameter needed for objective function of LC-KSVD dictionary learning algorithm is tried to achieve respectively.

The objective function of LC-KSVD dictionary learning algorithm is:

$<D,W,A,X>=\arg \underset{D,W,A,X}{\min }{\left|\right|Y-\mathrm{DX}\left|\right|}_2^2+{\mathrm{\&beta;}}_1{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+{\mathrm{\&beta;}}_2{\left|\right|H-\mathrm{WX}\left|\right|}_2^2;$

It meets $\&ForAll;k,{\left|\right|x_k\left|\right|}_0\&le;S_p;$

Wherein x _krepresent the kth sparse coefficient in the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection; S _prepresent degree of rarefication threshold value, in the methods of the invention, degree of rarefication threshold value S _ppreferred span is [1,20]; || || ₀for l ₀norm operational symbol, || || ₂for l ₂norm operational symbol; β ₁and β ₂represent sparse coefficient Error weight value and sorter Error weight value in objective function respectively, in the methods of the invention, described sparse coefficient Error weight value β ₁with sorter Error weight value β ₂preferred span be (0,1].

Can be learnt by the theories of learning of this objective function and LC-KSVD dictionary learning algorithm, the learning parameter needed for it comprises: original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀, initialization sorter matrix W ₀, label matrix H corresponding to training sample time-frequency data matrix vector Y, in order to represent training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between classification corresponding relation discriminating matrix Q and utilize original chemical handwriting practicing allusion quotation matrix D ₀training sample time-frequency data matrix vector Y is carried out to the sparse coefficient vector X of Its Sparse Decomposition gained.The concrete mode that this step asks for these learning parameters is:

41) respectively K-SVD dictionary learning is carried out to the time-frequency matrix of each training sample in each classification, obtain the KSVD dictionary that the training sample of each classification is corresponding, and respectively corresponding class label is marked to the dictionary atom in the KSVD dictionary of each classification, thus form original chemical handwriting practicing allusion quotation matrix D by each set with the dictionary atom of class label of all categories ₀;

42) original chemical handwriting practicing allusion quotation matrix D is utilized ₀, adopt sparse coding learning algorithm to carry out Its Sparse Decomposition to training sample time-frequency data matrix vector Y:

Y＝D ₀X；

Obtain the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding;

43) the label matrix H that training sample time-frequency data matrix vector Y is corresponding is generated; Described label matrix H=[h ₁, h ₂..., h _k..., h _k], k ∈ 1,2 ..., K}, K represent total number of training sample, h _kfor the column vector of k row in label matrix H, represent a kth training sample time-frequency data matrix y in training sample time-frequency data matrix vector Y _kclass label vector, class label vector h _kcomprise M numerical value, M represents the classification sum of the training sample of collection, wherein only training sample time-frequency data matrix y _kaffiliated m numerical value corresponding to m class is 1, and its remainder values is 0, i.e. h _k=[0 ..., 1 ..., 0] ^t, T is transposition symbol;

44) according to training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between classification corresponding relation, generate and differentiate matrix Q; Described discriminating matrix Q=[q ₁, q ₂..., q _n..., q _n], n ∈ 1,2 ..., N}, q _nrepresent the column vector differentiating n row in matrix Q, and k ∈ 1,2 ..., K}, wherein value in order to represent training sample time-frequency data matrix vector Y in a kth training sample time-frequency data matrix y _kwith original chemical handwriting practicing allusion quotation matrix D ₀in the n-th dictionary atom d _nclassification corresponding relation, if a described kth training sample time-frequency data matrix y _kwith described n-th dictionary atom d _nbelong to identical category, then get otherwise get k represents total number of training sample, and N represents original chemical handwriting practicing allusion quotation matrix D ₀in the total number of dictionary atom that comprises, T is transposition symbol;

45) pass through based on quadratic loss function and l ₂the polynary ridge regression equation of norm, solves respectively and obtains initialization linear transformation matrix A ₀with initialization sorter matrix W ₀:

A ₀＝(XX ^T+λ ₂I) ^-1XQ ^T；

W ₀＝(XX ^T+λ ₁I) ^-1XH ^T；

Wherein, I is the unit matrix of K × K, λ ₁and λ ₂for the specific gravity control parameter of polynary ridge regression equation, T is transposition symbol.

5) by original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀with initialization sorter matrix W ₀respectively as the initial value of LC-KSVD dictionary learning algorithm learning dictionary matrix D, linear transformation matrix A and sorter matrix W, and by training sample time-frequency data matrix corresponding sparse coefficient vector X, the label matrix H of vector Y with differentiate that matrix Q substitutes in the objective function of LC-KSVD dictionary learning algorithm in the lump and carry out learning and training, obtain the study dictionary matrix D after LC-KSVD dictionary learning _new, linear transformation matrix A _newwith sorter matrix W _new.

Based target function carries out the solution procedure of LC-KSVD dictionary learning, namely carry out according to the study treatment scheme of LC-KSVD dictionary learning algorithm, specifically can see existing document " Zhuolin Jiang; Zhe Lin and Larry S.Davis.Label Consistent K-SVD:Learning a Discriminative Dictionary for Recognition [J] .IEEE Transactions on Patttern Analysisand Machine Intelligence.2013,35 (11): 2651-2664 ".

6) study dictionary matrix D is utilized _new, linear transformation matrix A _newwith sorter matrix W _newcalculate benchmark dictionary matrix reference Transforming matrix with benchmark sorter matrix

$\hat{D}=\{\frac{d_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{d_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{d_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{d_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{A}=\{\frac{a_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{a_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{a_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{a_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{W}=\{\frac{w_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{w_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{w_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{w_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

Wherein, d _{new, k}represent study dictionary matrix D _newa kth dictionary atom, a _{new, k}represent linear transformation matrix A _newa kth column vector, w _{new, k}presentation class device matrix W _newa kth column vector, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection; || || ₂for l ₂norm operational symbol;

7) for radar target to be measured, gather the SAR image of radar target to be measured, obtain the time-frequency matrix z of radar target to be measured according to step 1 ~ 2 process, and utilize benchmark dictionary matrix adopt sparse coding learning algorithm that the time-frequency matrix z of radar target to be measured is carried out Its Sparse Decomposition:

$z=\hat{D}{x}_z;$

Obtain the sparse coefficient x that the time-frequency matrix z of radar target to be measured is corresponding _z;

8) benchmark sorter matrix is utilized determine the radar target classification j belonging to radar target to be measured:

$j=\arg \underset{j}{\max }(J=\hat{W}{x}_z);$

Wherein, J represents the class label vector that radar target to be measured is corresponding, and j ∈ J;

Realize the Classification and Identification to radar target to be measured thus.

Below by embodiment, technical scheme of the present invention is further described.

Embodiment:

The data image that the present embodiment utilizes MSTAR public database to announce, carrys out comparative evaluation and the present invention is based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook and the recognition effect of other Technology of Radar Target Identification.The present embodiment have chosen ten class radar targets that MSTAR public database the publishes data as experimental data base.This ten classes radar target is ground military vehicle or civilian vehicle, and outer shape has similarity, its radar target code name is respectively BMP2 (Infantry Tank), BRDM2 (amphibious armo(u)red scoutcar), BTR60 (armoring transfer cart), BTR70 (armored personnel carrier), D7 (agricultural dozer), T62 (T-62 type main website tank), T72 (T-72 type main website tank), ZIL131 (military trucks), ZSU234 (Self propelled Antiaircraft Gun battlebus) and 2S1 (carriage motor howitzer battlebus).Shown in the picture that the visible images of this ten classes radar target is as corresponding in respective code in Fig. 4 respectively.In MSTAR public database, store the some radar target images of this ten classes radar target in the different angle of pitch, different orientations shooting, the present embodiment therefrom have chosen the part radar target image of this ten classes radar target under the luffing angle of 17 ° and 15 ° captured by multiple different orientations and tests, wherein, using the training sample of the radar target image of 17 ° of angle of pitch shootings as experiment, the radar target image of 15 ° of angle of pitch shootings is made sample to be tested, in order to carry out radar target recognition test.Choose at the present embodiment in the experimental data base obtained, training sample quantity and the sample size to be tested of all kinds of radar target are as shown in table 1.

Table 1

Main information due to target distance image time-frequency figure all concentrates on the centre position of time-frequency figure, in order to reduce the data volume of time-frequency data processing further, reduce the requirement to computer memory in test, to save the time of sparse coding computing, the present embodiment is in experiment for target identification, be the data splitting 40 × 100 in the middle of it the former time-frequency figure of 100 × 100 from size, and in order to better utilize the detailed information in time-frequency figure, logarithm enhancing is carried out to the time-frequency diagram data after segmentation, thus split carrying out and logarithm strengthen process after the time-frequency diagram data that obtains test as the time audio data of training sample.Fig. 4 is the time-frequency diagram data after the former time-frequency figure of BMP2, BTR70 and T72 and the time-frequency figure after splitting and logarithm strengthen respectively; Wherein, figure (4a) one row are respectively the former time-frequency figure of BMP2, BTR70 and T72, figure (4b) one row are respectively BMP2, BTR70 and T72 time-frequency figure after over-segmentation, and figure (4c) row are respectively the time-frequency figure after to BMP2, BTR70 and T72 segmentation to carry out logarithm and strengthen the time-frequency diagram data after process.Through over-segmentation with after strengthening process, convert two-dimentional time-frequency diagram data to column vector, the flow process that then the present invention is based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook carries out SAR experiment for target identification respectively.

In the present embodiment, select BMP2, BTR70 and T72 tri-class vehicle target carry out experiment for target identification.In this tertiary target identification test of BMP2, BTR70 and T72, four kinds of algorithms all have employed identical training data.This training data is made up of 765 data under 17 ° of angles of pitch, comprising 233 data of SN-9563 model in BMP2, and 299 data of SN-132 model in 233 data of BTR70 and T72.And test data is made up of 1683 data under 15 ° of angles of pitch.Test data comprise respectively 195 data of SN-9563 model in BMP2, SN-9566 model 196 numbers according to this and 196 of SNC-21 model data, in 233 data of BTR70 and T72,274 numbers of 274 data of SN-132 model, SN-812 model are according to this and 274 of SNS-7 model data.In order to parameter in analytical algorithm is on the impact of recognition performance, carry out the experiment for target identification under different degree of rarefication threshold parameter and the experiment for target identification under different dictionary dimension respectively.In experiment, the method propose the present invention and the recognition result of the direct rarefaction representation method (SRC) of Range Profile time-frequency figure and Range Profile time-frequency figure K-SVD method compare analysis.SRC method comes from document " J.Wright; A.Yang; A.Ganesh and et al.Robust face recognition via sparse representation [J] .EEE Transactions on Patttern Analysisand Machine Intelligence.2009,31 (2): 210-227 ".K-SVD dictionary learning algorithm comes from document " M.Aharon; M.Elad and A.Bruckstein.K-SVD:An Algorithm for Designing Overcomplete Dictionaries for Sparse Representation [J] .IEEE Transactions on Signal Processing; 2006,54 (1): 4311-4322 ".

(1) the experiment for target identification result, under different degree of rarefication threshold parameter and analysis:

In this experiment, choose 1,5,10,15 degree of rarefication threshold parameters different with 20 5 kinds of values respectively and test, the dictionary dimension of four kinds of recognizers is all set to 500.This considers that SRC method directly adopts training data to form dictionary, and larger dictionary dimension can make test data obtain better rarefaction representation under SRC method.That is, each class target has 200 dictionary atoms.It should be noted that, because the amount of training data of this each class of tertiary target of BMP2, BTR70 and T72 is all greater than 200, in SRC algorithm, dictionary is obtained by 200 data of choosing random in each class training data herein.And K-SVD dictionary, LC-KSVD differentiates that dictionary is then utilize all training data study of tertiary target to obtain.In order to obtain objective experimental data, SRC discrimination result carries out the average recognition rate after the identification of test data respectively by the training data of 20 Stochastic choice.Fig. 5 adopts SRC (corresponding diagram (5a)), K-SVD (corresponding diagram (5b)), LC-KSVD (corresponding diagram (5c)) three kinds of algorithms under different degree of rarefication to the correct recognition rata curve of tertiary target.As seen from Figure 5, SRC algorithm, when degree of rarefication is 1, reaches best to the average recognition rate of tertiary target.Respectively 73.84%, 68.59% and 75.93% is reached to the correct recognition rata of BMP2, BTR70 and T72.K-SVD algorithm, when degree of rarefication is 1, reaches best to the average recognition rate of tertiary target.Respectively 68.95%, 70.01% and 75.76% is reached to the correct recognition rata of BMP2, BTR70 and T72.LC-KSVD differentiates that dictionary algorithm is when degree of rarefication is 10, and reaching best to the average recognition rate of tertiary target, is also that in four kinds of algorithms, average recognition rate is the highest.Respectively 94.74%, 97.95% and 95.13% is reached to the correct recognition rata of BMP2, BTR70 and T72.The best average recognition rate of three kinds of algorithms is as shown in table 2.Visible, the SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook that the present invention proposes all is being higher than SRC algorithm and K-SVD algorithm to the discrimination of tertiary target.

The best identified rate (%) of table 2 three kinds of algorithms under different degree of rarefication

Data in further analytical table 2 can find, SRC algorithm, when degree of rarefication is 1, is up to 72.78% to the average recognition rate of tertiary target; K-SVD dictionary learning algorithm, when degree of rarefication is 1, is up to 71.57% to the average recognition rate of tertiary target; And LC-KSVD differentiates that dictionary learning algorithm is when degree of rarefication is 10, is up to 95.94% to the average recognition rate of tertiary target.Therefore, when dictionary dimension is 500, in the identification experiment of tertiary target, the discrimination of the SAR target identification method based on the other dictionary learning of Range Profile time-frequency illustrated handbook that the present invention proposes is the highest.Trace it to its cause, just because of openness for objective function with the minimum and coefficient of reconstructed error in dictionary learning of K-SVD algorithm, cause it can not reach the requirement of high precision identification.SRC algorithm take directly training sample as dictionary, does not carry out choosing of feature to training sample.Therefore, the SRC method effect that can not obtain in object identification test.And differentiate that dictionary learning algorithm adds label consistent rule condition in the objective function function of dictionary learning, make by this Algorithm Learning to the sparse coding of dictionary to the test data of identical category have height similarity (show as sparse coding and there is identical class label after linear classifier effect), and what add again sparse coding in objective function can driscrimination error item, further enhancing the distinctive of sparse coding.Fig. 6 and Fig. 7 sets forth Range Profile time-frequency figure K-SVD dictionary learning and LC-KSVD differentiates the time-frequency figure dictionary that dictionary learning obtains.

(2) the experiment for target identification result, under different dictionary dimension and analysis:

In view of in the impact experiment of degree of rarefication on recognition result, SRC method and K-SVD dictionary learning algorithm are the highest to the average recognition rate of tertiary target when degree of rarefication is 1, and LC-KSVD differentiates that dictionary learning algorithm is the highest to the discrimination of tertiary target when degree of rarefication is 10.So, in the impact experiment of dictionary dimension on recognition result, carry out when degree of rarefication is 1 and 10 the experiment for target identification that dictionary dimension is 150,300,450 and 600 respectively.Experimental result as shown in Figure 8 and Figure 9, wherein, Fig. 8 be in the embodiment of the present invention when degree of rarefication threshold value value is 1 dictionary dimension to employing SRC (corresponding diagram (8a)), K-SVD (corresponding diagram (8b)), the influence curve comparison diagram of the discrimination of LC-KSVD (corresponding diagram (8c)) three kinds of algorithms, Fig. 9 be in the embodiment of the present invention when degree of rarefication threshold value value is 10 dictionary dimension to employing SRC (corresponding diagram (9a)), K-SVD (corresponding diagram (9b)), the influence curve comparison diagram of the discrimination of LC-KSVD (corresponding diagram (9c)) three kinds of algorithms.When dictionary dimension is 300, best average recognition rate is reached to BMP2, BTR70 and T72 from Fig. 9 and Fig. 9, SRC algorithm.K-SVD dictionary learning algorithm discrimination when dictionary size is 300 and 450 is lower, and when dictionary size is 600, discrimination is the highest.LC-KSVD differentiates that dictionary learning algorithm is when dictionary size is more than 300, and recognition effect is more stable, but is also that recognition effect is best when dictionary size is 600.Therefore, can think that these three kinds of algorithms have best recognition effect when dictionary size is 600.

In sum, above, the best identified rate of several algorithm to tertiary target is as shown in table 3.As shown in Table 3, the best average recognition rate of SRC algorithm to tertiary target is the best average recognition rate of 74.47%, K-SVD dictionary learning algorithm to tertiary target is 73.89%.LC-KSVD differentiates that the best average recognition rate of dictionary learning algorithm to tertiary target is 96.98%.

The best identified rate of table 3 three kinds of methods under different dictionary dimension

Carry out analysis to the result of above two groups of experiments further to find: under some degree of rarefication threshold parameter and dictionary dimension, dictionary learning is very high to a certain classification target discrimination.This phenomenon is relevant with concrete target.

Following table 4 gives best identified rate in the inventive method result of study and SRC, the best identified rate of K-SVD method and document " Zhao Q, Principe J C.Support vector machines for SAR automatic target recognition.IEEE Trans.on Aerospace and Electronic Systems.2001, 2 (37), 643 ~ 654 " SVM algorithm and document " J.J.Thiagarajan is adopted in, K.N.Ramamurthy, P.Knee, et al.Sparse representations for automatic target classification in SAR images [C] .Proceedings of the 4th International Symposium on Communications, Control and Signal Processing (ISCCSP), 2010:1-4 " in based on the discrimination of the SRC algorithm of two-dimensional SAR image.It should be noted that, the present invention and document " Zhao Q, Principe J C.Support vector machines for SAR automatic target recognition.IEEE Trans.on Aerospace and Electronic Systems.2001, 2 (37), 643 ~ 654 " and " J.J.Thiagarajan, K.N.Ramamurthy, P.Knee, et al.Sparse representations for automatic target classification in SAR images [C] .Proceedings of the 4th International Symposium on Communications, Control and Signal Processing (ISCCSP), 2010:1-4 " in adopt be identical test data.Difference is that the inventive method is the SAR target identification based on Range Profile time-frequency figure, and document " Zhao Q, Principe J C.Support vector machines for SAR automatic target recognition.IEEE Trans.on Aerospace and Electronic Systems.2001, 2 (37), 643 ~ 654 " the SVM method in and document " J.J.Thiagarajan, K.N.Ramamurthy, P.Knee, et al.Sparse representations for automatic target classification in SAR images [C] .Proceedings of the 4th International Symposium on Communications, Control and Signal Processing (ISCCSP), 2010:1-4 " in SRC algorithm be all SAR target identification based on two dimensional image.

Several distinct methods of table 4 compares (%) tertiary target best identified rate

As can be seen from Table 4, the average recognition rate of SAR target identification method to tertiary target that the present invention is based on the other dictionary learning of Range Profile time-frequency illustrated handbook reaches 96.98%; And the average recognition rate adopting SVM algorithm in document " Zhao Q; Principe J C.Support vector machines for SAR automatic target recognition.IEEE Trans.on Aerospace and Electronic Systems.2001; 2 (37), 643 ~ 654 " is 90.99%; Average recognition rate based on the SRC algorithm of image in document " J.J.Thiagarajan; K.N.Ramamurthy; P.Knee; et al.Sparse representations for automatic target classification in SAR images [C] .Proceedings of the 4th International Symposium on Communications; Control and Signal Processing (ISCCSP), 2010:1-4 " is 93.05%.Therefore, the SAR target identification method correct recognition rata in this several SAR Target Recognition Algorithms that the present invention is based on the other dictionary learning of Range Profile time-frequency illustrated handbook is the highest.This illustrates and is better than target recognition effect based on two-dimensional SAR image based on the SAR target recognition effect of Range Profile time-frequency figure, also show the validity of the inventive method.In addition, when training data all can not be included SRC dictionary too much in or limit dictionary dimension, how choosing training data for SRC algorithm is a problem required study.But concerning the present invention propose based on the other dictionary learning of Range Profile time-frequency illustrated handbook SAR target identification method, there is not this problem, this is also the place that the inventive method is more superior than SRC method.

In sum, the SAR target identification method that the present invention is based on the other dictionary learning of Range Profile time-frequency illustrated handbook is adopted to carry out radar target recognition, owing to have employed the Range Profile time-frequency figure of SAR as recognition feature, therefore avoid cause defocusing due to target travel or the factor such as signal to noise ratio (S/N ratio) causes picture quality not high time impact on target recognition effect, and carry out dictionary learning and sorter training by differentiating that dictionary learning is combined, characteristic information when effectively can extract Radar Target Using Range Profiles in audio data, not only be conducive to reducing the atom number in dictionary, reduce the computational complexity in sparse coding process, also help the precision improving sparse coding simultaneously, thus the recognition accuracy improved radar target, in addition, in the whole identifying of recognition methods of the present invention, all do not need to carry out position angle estimation to SAR image target, because this reducing identification complexity, it also avoid and identify that accuracy is to the dependence of target azimuth angular estimation, combine based on Range Profile time-frequency figure simultaneously and differentiate that dictionary learning technology is carried out radar target recognition and also had good recognition performance in a noisy environment, the robust performance promoting radar target recognition can be helped.

What finally illustrate is, above embodiment is only in order to illustrate technical scheme of the present invention and unrestricted, although with reference to preferred embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that, can modify to technical scheme of the present invention or equivalent replacement, and not departing from aim and the scope of technical solution of the present invention, it all should be encompassed in the middle of right of the present invention.

Claims (6)

1., based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, it is characterized in that, comprise the steps:

1) SAR image is converted to the Range Profile of SAR;

2) adopt adaptive Gaussian representation method, utilize Gaussian bases to carry out Breaking Recurrently expression to the Range Profile of SAR, and then calculate the time-frequency matrix of Range Profile of SAR;

3) for multiple different classes of known radar target, gather the SAR image of multiple known radar target respectively as training sample, and process the time-frequency matrix obtaining each training sample in each classification respectively according to step 1 ~ 2, thus arrange composing training sample time-frequency data matrix vector Y by the time-frequency matrix of each training sample of each classification;

4) based on training sample time-frequency data matrix vector Y and the classification information of each training sample, the learning parameter needed for objective function of LC-KSVD dictionary learning algorithm is tried to achieve respectively; Described learning parameter comprises original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀, initialization sorter matrix W ₀, label matrix H corresponding to training sample time-frequency data matrix vector Y, in order to represent training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between classification corresponding relation discriminating matrix Q and utilize original chemical handwriting practicing allusion quotation matrix D ₀training sample time-frequency data matrix vector Y is carried out to the sparse coefficient vector X of Its Sparse Decomposition gained;

5) by original chemical handwriting practicing allusion quotation matrix D ₀, initialization linear transformation matrix A ₀with initialization sorter matrix W ₀respectively as the initial value of LC-KSVD dictionary learning algorithm learning dictionary matrix D, linear transformation matrix A and sorter matrix W, and by training sample time-frequency data matrix corresponding sparse coefficient vector X, the label matrix H of vector Y with differentiate that matrix Q substitutes in the objective function of LC-KSVD dictionary learning algorithm in the lump and carry out learning and training, obtain the study dictionary matrix D after LC-KSVD dictionary learning _new, linear transformation matrix A _newwith sorter matrix W _new;

6) study dictionary matrix D is utilized _new, linear transformation matrix A _newwith sorter matrix W _newcalculate benchmark dictionary matrix reference Transforming matrix with benchmark sorter matrix

$\hat{D}=\{\frac{d_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{d_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{d_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{d_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{A}=\{\frac{a_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{a_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2},...,\frac{a_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{a_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

$\hat{W}=\{\frac{w_{\mathrm{new},1}}{{\left|\right|d_{\mathrm{new},1}\left|\right|}_2},\frac{w_{\mathrm{new},2}}{{\left|\right|d_{\mathrm{new},2}\left|\right|}_2}...,\frac{w_{\mathrm{new},k}}{{\left|\right|d_{\mathrm{new},k}\left|\right|}_2},...,\frac{w_{\mathrm{new},K}}{{\left|\right|d_{\mathrm{new},K}\left|\right|}_2}\};$

Wherein, d _{new, k}represent study dictionary matrix D _newa kth dictionary atom, a _{new, k}represent linear transformation matrix A _newa kth column vector, w _{new, k}presentation class device matrix W _newa kth column vector, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection; || || ₂for l ₂norm operational symbol;

7) for radar target to be measured, gather the SAR image of radar target to be measured, obtain the time-frequency matrix z of radar target to be measured according to step 1 ~ 2 process, and utilize benchmark dictionary matrix adopt sparse coding learning algorithm that the time-frequency matrix z of radar target to be measured is carried out Its Sparse Decomposition:

$z={\hat{D}x}_z;$

Obtain the sparse coefficient x that the time-frequency matrix z of radar target to be measured is corresponding _z;

8) benchmark sorter matrix is utilized determine the radar target classification j belonging to radar target to be measured:

$j=\arg \underset{j}{\max }(J={\hat{W}x}_z);$

Wherein, J represents the class label vector that radar target to be measured is corresponding, and j ∈ J;

Realize the Classification and Identification to radar target to be measured thus.

2., according to claim 1 based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, it is characterized in that, described step 2 is specially:

Adopt adaptive Gaussian representation method, utilize Gaussian bases to carry out Breaking Recurrently expression to the Range Profile of SAR:

$r\left(t\right)=\underset{i=1}{\overset{i_{\max }}{\mathrm{\&Sigma;}}}{C}_i{g}_i\left(t\right)+r_{i_{\max }+1}\left(t\right),$ And ${\left|\right|r\left(t\right)\left|\right|}^2=\underset{i=0}{\overset{i_{\max }}{\mathrm{\&Sigma;}}}{\left|C_i\right|}^2+{\left|\right|r_{i_{\max }+1}\left(t\right)\left|\right|}^2;$

Wherein, t represents the time; R (t) represents the Range Profile of the SAR of time t; g _it () represents that picture r (t) of adjusting the distance carries out the Gaussian bases of i-th Breaking Recurrently, C _irepresent that picture r (t) of adjusting the distance carries out the expansion coefficient of i-th Breaking Recurrently, i ∈ 1,2 ..., i _max, i _maxrepresent the iteration total degree carrying out Breaking Recurrently expression, and iteration total degree i _maxmake the reconstructed error r of Breaking Recurrently _imax+1t () meets ε is default reconstructed error threshold value, and 10 ^-3≤ ε≤10 ^-5; And have relational expression:

$g_i\left(t\right)={\left(\frac{1}{{\mathrm{\&pi;\&alpha;}}_i}\right)}^{\frac{1}{4}}\exp \{-\frac{{(t-t_i)}^2}{{2\mathrm{\&alpha;}}_i}\}\&CenterDot;\exp \left(j2\mathrm{\&pi;}{f}_{i}t\right);$

${\left|C_i\right|}^2=\underset{t_i,f_i,{\mathrm{\&alpha;}}_i}{\max }{|\&Integral;r_i\left(t\right)g_i^{*}\left(t\right)\mathrm{dt}|}^2,{\mathrm{\&alpha;}}_i\&Element;R^{+},t_i,f_i\&Element;R;$

T _i, f _i, α _irepresent that picture r (t) of adjusting the distance carries out the resolution parameter in the Gaussian bases of i-th Breaking Recurrently; r _it () represents that picture r (t) of adjusting the distance carries out the iteration discrepance before i-th Breaking Recurrently, and $r_i\left(t\right)=\left\{\begin{array}{cc}r\left(t\right),& i=1\\ r_{i-1}\left(t\right)-C_{i-1}{g}_{i-1}\left(t\right),& i\&GreaterEqual;2\end{array}\right.;$ represent Gaussian bases g _ithe conjugation of (t); || represent the operational symbol that takes absolute value; represent adjustment resolution parameter t _i, f _i, α _imake || ²maximum maximum operator;

Adopt Fourier Transform Algorithm to solve above-mentioned relation formula, obtain the resolution parameter t of i-th Breaking Recurrently _i, f _i, α _iwith expansion coefficient C _i, and then time-frequency matrix Θ (t, f) of Range Profile r (t) is obtained by following formula:

$\mathrm{\&Theta;}(t,f)=\underset{i=0}{\overset{i_{\max }}{\mathrm{\&Sigma;}}}2{\left|C_i\right|}^2\exp \{-\frac{{(t-t_i)}^2}{{2\mathrm{\&alpha;}}_i}-{\left(2\mathrm{\&pi;}\right)}^2{\mathrm{\&alpha;}}_i{\left({f-f}_i\right)}^2\};$

Wherein, t represents the time, and f represents frequency.

3., according to claim 1 based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, it is characterized in that, described step 4 is specially:

41) respectively K-SVD dictionary learning is carried out to the time-frequency matrix of each training sample in each classification, obtain the KSVD dictionary that the training sample of each classification is corresponding, and respectively corresponding class label is marked to the dictionary atom in the KSVD dictionary of each classification, thus form original chemical handwriting practicing allusion quotation matrix D by each set with the dictionary atom of class label of all categories ₀;

42) original chemical handwriting practicing allusion quotation matrix D is utilized ₀, adopt sparse coding learning algorithm to carry out Its Sparse Decomposition to training sample time-frequency data matrix vector Y:

Y＝D ₀X；

Obtain the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding;

43) the label matrix H that training sample time-frequency data matrix vector Y is corresponding is generated; Described label matrix H=[h ₁, h ₂..., h _k..., h _k], k ∈ 1,2 ..., K}, K represent total number of training sample, h _kfor the column vector of k row in label matrix H, represent a kth training sample time-frequency data matrix y in training sample time-frequency data matrix vector Y _kclass label vector, class label vector h _kcomprise M numerical value, M represents the classification sum of the training sample of collection, wherein only training sample time-frequency data matrix y _kaffiliated m numerical value corresponding to m class is 1, and its remainder values is 0, i.e. h _k=[0 ..., 1 ..., 0] ^t, T is transposition symbol;

44) according to training sample time-frequency data matrix vector Y and original chemical handwriting practicing allusion quotation matrix D ₀between classification corresponding relation, generate and differentiate matrix Q; Described discriminating matrix Q=[q ₁, q ₂..., q _n..., q _n], n ∈ 1,2 ..., N}, q _nrepresent the column vector differentiating n row in matrix Q, and k ∈ 1,2 ..., K}, wherein value in order to represent training sample time-frequency data matrix vector Y in a kth training sample time-frequency data matrix y _kwith original chemical handwriting practicing allusion quotation matrix D ₀in the n-th dictionary atom d _nclassification corresponding relation, if a described kth training sample time-frequency data matrix y _kwith described n-th dictionary atom d _nbelong to identical category, then get otherwise get k represents total number of training sample, and N represents original chemical handwriting practicing allusion quotation matrix D ₀in the total number of dictionary atom that comprises, T is transposition symbol;

45) pass through based on quadratic loss function and l ₂the polynary ridge regression equation of norm, solves respectively and obtains initialization linear transformation matrix A ₀with initialization sorter matrix W ₀:

A ₀＝(XX ^T+λ ₂I) ^-1XQ ^T；

W ₀＝(XX ^T+λ ₁I) ^-1XH ^T；

Wherein, I is the unit matrix of K × K, λ ₁and λ ₂for the specific gravity control parameter of polynary ridge regression equation, T is transposition symbol.

4. according to claim 1 based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, it is characterized in that, the objective function of described LC-KSVD dictionary learning algorithm is:

$<D,W,A,X>=\arg \underset{D,W,A,X}{\min }{\left|\right|Y-\mathrm{DX}\left|\right|}_2^2+{\mathrm{\&beta;}}_1{\left|\right|Q-\mathrm{AX}\left|\right|}_2^2+{\mathrm{\&beta;}}_2{\left|\right|H-\mathrm{WX}\left|\right|}_2^2;$

It meets $\&ForAll;k,{\left|\right|x_k\left|\right|}_0\&le;S_p;$

Wherein, x _krepresent the kth sparse coefficient in the sparse coefficient vector X that training sample time-frequency data matrix vector Y is corresponding, k ∈ 1,2 ..., K}, K represent total number of the training sample of collection; S _prepresent degree of rarefication threshold value; || || ₀for l ₀norm operational symbol, || || ₂for l ₂norm operational symbol; β ₁and β ₂represent sparse coefficient Error weight value and sorter Error weight value in objective function respectively.

5. according to claim 4 based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, it is characterized in that, described degree of rarefication threshold value S _pspan be [1,20].

6. according to claim 4 based on the SAR target identification method of the other dictionary learning of Range Profile time-frequency illustrated handbook, it is characterized in that, described sparse coefficient Error weight value β ₁with sorter Error weight value β ₂span be (0,1].