CN104021399B - SAR object identification method based on range profile time-frequency diagram non-negative sparse coding - Google Patents

Abstract

The invention provides an SAR object identification method based on range profile time-frequency diagram non-negative sparse coding. The method utilizes non-negative sparse coding, and in the whole identification process, SAR image objects do not need azimuth angle estimation, thereby reducing the identification complex degree, avoiding the dependence of identification accuracy on object azimuth angle estimation, and helping to improve the object identification rate; and meanwhile, radar object identification is carried out based on the range profile time-frequency diagram non-negative sparse coding technique, and good identification performance is also achieved in the noise environment, thereby not influencing identification effect when not-high image quality, due to factors of defocusing or signal to noise ratio and the like during the object moving, is caused, and helping to improve the robustness performance of the radar object identification.

Description

SAR target identification methods based on Range Profile time-frequency figure non-negative sparse coding

Technical field

It is the present invention relates to Technology of Radar Target Identification field, more particularly to a kind of based on Range Profile time-frequency figure non-negative sparse volume The SAR target identification methods of code.

Background technology

Synthetic aperture radar（Synthetic Aperture Radar, abbreviation SAR）Technology, be using be mounted in satellite or Movable radar on aircraft, obtains a kind of pulse radar technology of the geographical band radar target image of high accuracy.Due to the master of SAR Complicated scattering mechanism in dynamic imaging characteristicses and imaging process, the target property and optical imagery difference in SAR image are very big, this Many difficulties are brought to target's feature-extraction and identification.

Scientific research personnel has studied many Target Recognition Algorithms based on two-dimensional SAR image.Wherein, most direct one The method of kind is exactly directly, as feature, to carry out the identification of target using SAR image.Another kind of radar target identification method is based on little Wave conversion or multiscale analysis.In addition, the characteristics of image such as target area description, shade is also used for carrying out target knowledge Not.Can provide a kind of fine using the feature based on physics of scattering center model, the related goal description of physics, but Which is needed by estimation azimuth of target, therefore the accuracy rate of its radar target recognition is also by the accurate of target bearing angular estimation Property limit, therefore be extremely difficult to very good recognition performance.

However, SAR target recognitions can also be using the method based on target distance image, target range seems one-dimensional data Image, is to obtain the SAR complex patterns of target through a series of process, obtains the Range Profile of SAR by SAR image conversion process Concrete grammar referring to document " 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”.Know with the radar target based on image Other method is compared, and can be extracted target-sensor orientation based on the advantage of the radar target identification method of Range Profile and be relied on Characteristic information.In addition, when the uncooperative motion due to target or as the low factor of signal to noise ratio causes target image fuzzy etc. When, the feature extraction based on two dimensional image and identification are difficult to prove effective, and the feature extraction based on Range Profile and identification are by contrast more It is advantageous.In terms of the SAR image target recognition based on target distance image, also there are some correlational studyes.In some documents SAR target recognitions are carried out using the One-dimensional scattering centres feature of Range Profile, the shortcoming of this method is：Point scattering can only be extracted special Levy, and the target complex electromagnetic feature such as extended distance, frequency dispersion, resonance can not be extracted, so as to recognize accuracy It is limited.Also research worker carries out SAR target recognitions using the high-order spectrum signature of Range Profile, and the shortcoming of this method is to adjust the distance The wavefront vibration sensing of picture, feature are not sufficiently stable, thus its recognition methods accuracy and robustness be all greatly affected.From As can be seen that how to extract the effective Range Profile feature of robust in these research work, it is the key of SAR target identification technologies.

The content of the invention

For the above-mentioned problems in the prior art, it is required for solve SAR image target recognition in prior art Estimate azimuth of target, the problem of identification limited accuracy, the invention provides a kind of be based on Range Profile time-frequency figure non-negative sparse The SAR target identification methods of coding, the radar target identification method are represented come to non-negative time-frequency plane data using non-negative sparse It is modeled and feature extraction, it is not necessary to which target bearing angular estimation is carried out to SAR image, while can avoid defocusing or noise Than etc. impact of the factor to target recognition effect, improve SAR target recognitions accuracy.

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

Based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding, comprise the steps：

A）SAR image is converted to into the Range Profile of SAR；

B）Using adaptive Gaussian representation method, breakdown is iterated to the Range Profile of SAR using Gaussian bases Show, and then be calculated the time-frequency matrix of the Range Profile of SAR；

C）For the different known radar target of multiclass, the SAR image of multiple known radar targets is gathered respectively as instruction Practice sample, and process the time-frequency matrix for obtaining each training sample in each classification according to step A～B respectively；Again each is instructed The all non-negative pixel arrangements practiced in the time-frequency matrix of sample form a column vector, so as to the time-frequency square by each training sample One training sample matrix Z of column vector arrangement form that battle array is constituted；Using non-negative sparse coding learning algorithm by training sample square Battle array Z is decomposed into the product of non-negative dictionary matrix D and non-negative sparse coefficient matrix H：

Z=DH；

D）The non-negative dictionary matrix obtained using training sample matrix decomposition, according to the time-frequency matrix meter of each training sample Calculation obtains the respective non-negative sparse coding characteristic vector of each training sample, so as to the non-negative for obtaining all kinds of known radar targets is dilute Dredge coding total characteristic vector；

E）For radar target to be measured, the SAR image of radar target to be measured is gathered, obtains to be measured according to step A～B process The time-frequency matrix of radar target；The non-negative dictionary matrix obtained using training sample matrix decomposition, according to radar target to be measured Time-frequency matrix calculus obtain the non-negative sparse coding characteristic vector of radar target to be measured；

F）It is vectorial as identification benchmark using the non-negative sparse coding total characteristic of all kinds of known radar targets, using support vector Machine recognizer, carries out Classification and Identification to the non-negative sparse coding characteristic vector of radar target to be measured, obtains radar target to be measured Recognition result.

It is in the above-mentioned SAR target identification methods based on Range Profile time-frequency figure non-negative sparse coding, preferably, described Step B is specially：

Using adaptive Gaussian representation method, exploded representation is iterated to the Range Profile of SAR using Gaussian bases：

And

Wherein, t express times；The Range Profile of the SAR of r (t) express time ts；g_iT () represents to adjust the distance and enters as r (t) The Gaussian bases that row ith iteration is decomposed, C_iExpression to be adjusted the distance and carry out the expansion coefficient of ith iteration decomposition, i ∈ as r (t) {1,2,…,i_max, i_maxExpression is iterated the iteration total degree of exploded representation, and iteration total degree i_maxSo that Breaking Recurrently Reconstructed errorMeetε is default reconstructed error threshold value, and 10^-3≤ε≤10^-5；And have relation Formula：

t_i,f_i,α_iExpression to be adjusted the distance and carry out the resolution parameter in the Gaussian bases of ith iteration decomposition as r (t)；r_i T () represents to adjust the distance and carry out the iteration discrepance before ith iteration decomposition as r (t), and Represent the conjugation of Gaussian bases gi (t)；| | represent the operator that takes absolute value；Represent adjustment resolution parameter t_i,f_i,α_iSo that | |²Maximum maximum operator；

Above-mentioned relation formula is solved using Fourier Transform Algorithm, the resolution parameter t of ith iteration decomposition is obtained_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：

Wherein, t express times, f represent frequency.

It is in the above-mentioned SAR target identification methods based on Range Profile time-frequency figure non-negative sparse coding, preferably, described Step C is specially：

c1）For the different known radar target of multiclass, the SAR image of multiple known radar targets is gathered respectively as instruction Practice sample, and process the time-frequency matrix for obtaining each training sample in each classification according to step A～B respectively；

c2）All N number of non-negative pixel arrangement in the time-frequency matrix of each training sample is formed into a column vector, so as to The training sample matrix Z of the one N × M of column vector arrangement form being made up of the time-frequency matrix of each training sample；M represents training The total number of sample；

c3）Using non-negative sparse coding learning algorithm by training sample matrix Z be decomposed into N × K non-negative dictionary matrix D and The product of the non-negative sparse coefficient matrix H of K × M：

Z=DH；

K represents the quantity of non-negative dictionary atom in non-negative sparse coding.

It is in the above-mentioned SAR target identification methods based on Range Profile time-frequency figure non-negative sparse coding, preferably, described Step D is specially：

For the time-frequency matrix of j-th training sample of pth class known radar targetUsing training sample matrix decomposition The non-negative dictionary matrix D for obtaining, calculates its corresponding non-negative sparse coding characteristic vector

(i)^-1Representing matrix inversion operation is accorded with；So as to all training samples by pth class known radar target are corresponding non- Negative sparse coding characteristic vector constitutes the non-negative sparse coding total characteristic vector V of pth class known radar target_p：

Wherein, J represents the training sample sum of pth class known radar target；Thus, all kinds of known radar mesh are respectively obtained Target non-negative sparse coding total characteristic vector.

It is in the above-mentioned SAR target identification methods based on Range Profile time-frequency figure non-negative sparse coding, preferably, described Step E is specially：

The SAR image of radar target to be measured is gathered, and the time-frequency matrix of radar target to be measured is obtained according to step A～B process z_x, the non-negative dictionary matrix D for then being obtained using training sample matrix decomposition, the corresponding non-negative sparse of calculating radar target to be measured Coding characteristic vector v_x：

v_x=(D^TD)^-1D^Tz_x；

(i)^-1Representing matrix inversion operation is accorded with.

Compared to prior art, the present invention has the advantages that：

1st, when causing to defocus or when the factor such as signal to noise ratio causes picture quality not high due to target motion, based on target The target recognition effect on driving birds is not good that SAR image is extracted, but SAR target knowledge of the present invention based on Range Profile time-frequency figure non-negative sparse coding Other method, then be not limited thereto.

2nd, SAR target identification method of the present invention based on Range Profile time-frequency figure non-negative sparse coding is compiled using non-negative sparse Code, can effectively model target distance image time-frequency characteristics, be favorably improved object recognition rate.

3rd, whole identification process of the present invention based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding In, orientation angular estimation need not be carried out to SAR image target, therefore reduces identification complexity, it also avoid identification accurate Really dependence of the property to target bearing angular estimation, while carry out radar target based on the non-negative sparse coding technology of Range Profile time-frequency figure Identification also has good recognition performance in a noisy environment, can help lift the robust performance of radar target recognition.

Description of the drawings

Fig. 1 is flow chart of the present invention based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding.

Exemplary plots of the Fig. 2 for the Range Profile of SAR.

Time-frequency matrix exemplary plots of the Fig. 3 for the Range Profile of SAR.

Fig. 4 be MSTAR public databases in code name be respectively BMP2, BRDM2, BTR60, BTR70, D7, T62, T72, The visible images of the ten class radar targets of ZIL131, ZSU234 and 2S1.

Fig. 5 be the embodiment of the present invention in dimension be 100 non-negative dictionary matrix.

Fig. 6 is for three kinds of radar target identification methods in the embodiment of the present invention to 10 classification target recognition effects with dictionary dimension The curve chart of change.

Fig. 7 is three kinds of radar target identification methods in the embodiment of the present invention with signal to noise ratio（SNR）The model from -5db to 20db Enclose the change curve of the correct identification ratio of change.

Specific embodiment

With reference to the accompanying drawings and examples technical scheme is further described.

The present invention proposes a kind of method of time-frequency domain non-negative sparse coding for extracting Radar Target Using Range Profiles feature, and For SAR target recognitions.Sparse coding is a kind of effective information representation model, has been successfully applied to image procossing, pattern The fields such as identification.In traditional rarefaction representation, data are described as the combination of basic feature, and these basic features include additivity With the composition of subtracting property.However, for non-negative data, subtracting property feature does not simultaneously meet the concept that local constitutes entirety.Based on this, occur Non-negative sparse coding, in this non-negative sparse representational framework, all of characteristic component is additivity, not subtracting property, It just, will not be negative that i.e. rarefaction representation coefficient is only.In the present invention, to target distance image use non-negative time frequency analysis, i.e., away from Time-frequency plane data from picture are also non-negative, therefore, represent to build non-negative time-frequency plane data using non-negative sparse Mould and feature extraction.Based on such mentality of designing, SAR target of the present invention based on Range Profile time-frequency figure non-negative sparse coding is known In other method, using non-negative sparse coding technology, low-dimensional of the Range Profile of SAR targets in time-frequency matrix is extracted and models sparse Local feature, and using these sparse local features as identification feature, using SVM recognizer to radar mesh to be measured Mark carries out Classification and Identification, even if in the case of non-negative dictionary matrix dimension is less remaining to obtain higher correct recognition rata, from And need to estimate azimuth of target, recognize asking for limited accuracy based on the SAR target recognitions of Range Profile in solving prior art Topic, reaches the purpose for improving SAR target recognition accuracys.

Overall flow such as Fig. 1 institute of the present invention based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding Show, specifically include following steps：

A）SAR image is converted to into the Range Profile of SAR.

Range Profile is one-dimensional data image, is to obtain the SAR complex patterns of target through a series of process, by SAR Image conversion process obtain the Range Profile of SAR concrete grammar be prior art, 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):There is more detailed introduction in 1230-1242 ", the present invention is repeated no more.What conversion process was obtained The example of the Range Profile of SAR is as shown in Figure 2.

B）Using adaptive Gaussian representation method, breakdown is iterated to the Range Profile of SAR using Gaussian bases Show, and then be calculated the time-frequency matrix of the Range Profile of SAR.

In prior art group, there are the time frequency analyzing tool of many signals, such as Short Time Fourier Transform, Wei Ge Receive-Willie distribution（Wigner-ville distribution, are abbreviated as WVD）, wavelet analysises, adaptive Gaussian representation （Adaptive Gaussian representation, are abbreviated as AGR）Deng.But, not all time-frequency analysis technology It is adapted to the feature extraction of Radar Target Using Range Profiles.The present invention selects AGR adaptive Gaussian representation method methods to carry out The time frequency analysis of Range Profile.Its reason is, compared to other Time-Frequency Analysis Method, in the time-frequency distributions local that AGR is obtained The heart can be corresponded to and the scattering mechanism phenomenon such as scattering center and local resonance just；Meanwhile, AGR can provide the joint of Range Profile Non-negative time-frequency distributions, and be adaptive, and no cross term interference.

The step is specially：

Using adaptive Gaussian representation method, exploded representation is iterated to the Range Profile of SAR using Gaussian bases：

And

Wherein, t express times；The Range Profile of the SAR of r (t) express time ts；g_iT () represents to adjust the distance and enters as r (t) The Gaussian bases that row ith iteration is decomposed, C_iExpression to be adjusted the distance and carry out the expansion coefficient of ith iteration decomposition, i ∈ as r (t) {1,2,…,i_max, i_maxExpression is iterated the iteration total degree of exploded representation, and iteration total degree i_maxSo that Breaking Recurrently Reconstructed errorMeetε is default reconstructed error threshold value, and 10^-3≤ε≤10^-5；And have relation Formula：

t_i,f_i,α_iExpression to be adjusted the distance and carry out the resolution parameter in the Gaussian bases of ith iteration decomposition as r (t)；r_i T () represents to adjust the distance and carry out the iteration discrepance before ith iteration decomposition as r (t), and have：

Represent Gaussian bases g_iThe conjugation of (t)；| | represent the operator that takes absolute value；Represent adjustment point Solution parameter t_i,f_i,α_iSo that | |²Maximum maximum operator；

Above-mentioned relation formula is solved using Fourier Transform Algorithm, the resolution parameter t of ith iteration decomposition is obtained_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：

Wherein, t express times, f represent frequency.

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

C）For the different known radar target of multiclass, the SAR image of multiple known radar targets is gathered respectively as instruction Practice sample, and process the time-frequency matrix for obtaining each training sample in each classification according to step A～B respectively；Again each is instructed The all non-negative pixel arrangements practiced in the time-frequency matrix of sample form a column vector, so as to the time-frequency square by each training sample One training sample matrix Z of column vector arrangement form that battle array is constituted；Using non-negative sparse coding learning algorithm by training sample square Battle array Z is decomposed into the product of non-negative dictionary matrix D and non-negative sparse coefficient matrix H：

Z=DH。

Non-negative sparse coding（Non-negative sparse coding, are abbreviated as NNSC）A kind of statistics of data from Adapt to method for expressing, there is provided a kind of important method of image is represented with training data.Here, because the time-frequency square of Range Profile Battle array is non-negative, therefore the time-frequency matrix of Range Profile can be regarded as image and processed.The basic thought of NNSC is：For tool There is the non-negative image z of n pixel, it is modeled as a series of linear combination of non-negative dictionary atoms, these non-negative dictionary atom tables It is shown as D₁,…,D_K, have these non-negative dictionary atomic building non-negative dictionary matrix D=[D₁,…,D_K], during K is non-negative sparse coding The quantity of non-negative dictionary atom；Therefore, image z is expressed as：

Wherein, each column vector of D is a non-negative dictionary atom, and each non-negative dictionary atom is a n dimension Vector.Therefore, D is a n × K matrix；K dimension no negative coefficient vector h give the contribution of each atom；The important hypothesis of NNSC Be coefficient vector h be sparse, its target is one non-negative dictionary matrix of design, expression that can be accurate and sparse in order to z. It has been proved that due to Condition of Non-Negative Constrains, NNSC is a kind of expression based on part, because only adding in the framework of sparse representation Property is without subtracting property atom；There was only minority non-negative dictionary atom in the openness solution that can ensure that the problems referred to above of coefficient vector It is movable effective；Such design can more select those represent the non-negative dictionary atom of Range Profile time-frequency distributions characteristic.And For Range Profile time-frequency matrix character is extracted, then can extend to represent a series of time-frequency matrix using non-negative sparse coding.

Based on this thought, the step is specially：

c1）For the different known radar target of multiclass, the SAR image of multiple known radar targets is gathered respectively as instruction Practice sample, and process the time-frequency matrix for obtaining each training sample in each classification according to step A～B respectively；

c2）All N number of non-negative pixel arrangement in the time-frequency matrix of each training sample is formed into a column vector；Due to The dimension of the time-frequency matrix of each training sample is related to the data points of training sample SAR Range Profiles, if training sample The data points of SAR Range Profiles are n, then the dimension of the time-frequency matrix of training sample as n × n's, therefore each training sample Non-negative pixel number in time-frequency matrix is N=n × n；So as to the column vector row being made up of the time-frequency matrix of each training sample Row form the training sample matrix Z of a N × M；M represents the total number of training sample；

c3）Using non-negative sparse coding learning algorithm by training sample matrix Z be decomposed into N × K non-negative dictionary matrix D and The product of the non-negative sparse coefficient matrix H of K × M：

Z=DH；

K represents the quantity of non-negative dictionary atom in non-negative sparse coding.

Non-negative sparse coding learning algorithm is existing algorithm known.Its object function is：

Its constraints is：For arbitrary n, m, k, there is D_nk≥0,H_km>=0, and the kth of non-negative dictionary matrix D Column vector d_kMeet | d_k|=1.Sparse factor lambda is a constant, and the constant controls the compromise between Accurate Reconstruction and degree of rarefication.ξ Formal definition degree of rarefication how to estimate, in general, select ξ (Hk_j)=|H_kj|.Learn with regard to non-negative sparse coding Object function, can be found in existing document " Zou F, Feng H, Ling H, et al.Nonnegative sparse coding induced hashing for image copy detection[J].Neurocomputing,2012,53,45-56”。

There are two parameters in NNSC learning algorithms：Sparse factor lambda and iteration step length μ；Object function is updated in following iteration It is convergence under rule：

D←D^T-μ(D^TH-Z)H^T；

Above-mentioned arrow represents that the fallback relationship that iteration updates, i.e., in next step iterative process, by the expression on the right of arrow Formula is replaced the data volume on the arrow left side and is iterated calculation process.Wherein []_kmThe multiplication and division of representing matrix are to element by element 's.With regard to the iterative convergent process of non-negative sparse coding object function, existing document " Hoyer P O, Modeling is can be found in receptive fields with non-negative sparse coding.Neurocomputing.2003,52,547- 552”。

D）The non-negative dictionary matrix obtained using training sample matrix decomposition, according to the time-frequency matrix meter of each training sample Calculation obtains the respective non-negative sparse coding characteristic vector of each training sample, so as to the non-negative for obtaining all kinds of known radar targets is dilute Dredge coding total characteristic vector.

It is for the identification process of SAR targets, main to include in terms of two.On the one hand it is training step, carries in training step The NNSC features of SAR target training datas are taken, these features can be used in follow-up identification step.It is test step on the other hand Suddenly, the SAR test datas of radar target to be measured are input in system, extract the NNSC features of test data.Then using knowledge Other algorithm is determining target type.

In the training step of the inventive method, for the time-frequency matrix of j-th training sample of pth class known radar targetThe non-negative dictionary matrix D obtained using training sample matrix decomposition, calculates its corresponding non-negative sparse coding characteristic vector

(i)^-1Representing matrix inversion operation is accorded with；So as to all training samples by pth class known radar target are corresponding non- Negative sparse coding characteristic vector constitutes the non-negative sparse coding total characteristic vector V of pth class known radar target_p：

Wherein, J represents the training sample sum of pth class known radar target；Thus, all kinds of known radar mesh are respectively obtained Target non-negative sparse coding total characteristic vector.

E）For radar target to be measured, the SAR image of radar target to be measured is gathered, obtains to be measured according to step A～B process The time-frequency matrix of radar target；The non-negative dictionary matrix obtained using training sample matrix decomposition, according to radar target to be measured Time-frequency matrix calculus obtain the non-negative sparse coding characteristic vector of radar target to be measured.

Specifically, the SAR image of collection radar target to be measured, according to step A～B process obtain radar target to be measured when Frequency matrix z_x, the non-negative dictionary matrix D for then being obtained using training sample matrix decomposition, calculating radar target to be measured are corresponding non- Negative sparse coding characteristic vector v_x：

v_x=(D^TD)^-1D^Tz_x；

(i)^-1Representing matrix inversion operation is accorded with.

F）It is vectorial as identification benchmark using the non-negative sparse coding total characteristic of all kinds of known radar targets, using support vector Machine recognizer, carries out Classification and Identification to the non-negative sparse coding characteristic vector of radar target to be measured, obtains radar target to be measured Recognition result.

It is achieved in the identification to radar target to be measured.

SVM（Support vector machine, referred to as SVM）Recognizer is also should in prior art With more ripe sorting technique.SVM classifier is the principle based on Structural risk minization.Assume that N number of training sample is designated as {v_i,y_i}_i=1,…,N, the result for training SVM is exactly hyperplane decision function：

Wherein, variable α_iLagrange multiplier operator, φ (v_i, v) be testing feature vector and training feature vector core letter Number, b is the skew of feature space.The classification of test data is that, according to decision function, it is super flat that it shows that this pattern would is that Which face in face.For φ (v_i, v), Radial basis kernel function is selected here：

Based on svm classifier method, for the present invention, it is preferred to use a pair in support vector cassification recognition methodss Multi-identification method.For the set V=V of the non-negative sparse coding feature sum vector of known k classes training sample₁∪…∪V_k, V_i ∈ V represent the non-negative sparse coding total characteristic vector of the i-th class known radar target.If the corresponding non-negative of radar target to be measured is dilute Thin coding characteristic vector v_x∈V_i, then radar target to be measured belong to the i-th class, and the type code symbol y=i that recognition result is obtained. According to One-Against-All recognition methodss（A pair of multi-identification methods）, then equivalent to the classification and identification a pair of k classes K man-to-man classification and identification is decomposed into, that is, has k decision function.OrderRepresent total sample number, wherein l_i Represent V_iNumber of samples in set.I-th classification and identification is to belong to V with SVM handle_iThe sample of set and do not belong to In V_iThe sample of set separates, that is, solve following Dual Programming Problem：

Maximize function：

Constraints is：

Wherein, αⁱFor the Lagrange multiplier operator of i-th liang of class problem, order：

Then the decision function of i-th man-to-man classification and identification is：

bⁱFor the skew of the feature space of i-th man-to-man classification and identification.Radar target to be measured is being carried out point When class is recognized, all k decision function f are calculated_i(v), i=1 ..., k.If only one of which f_i(v)>0, then radar target to be measured Belong to V_iClass.If more than one f_i(v)>0 or all k f_iV ()≤0, then the ownership of radar target to be measured is uncertain 's.At this moment using following judgement：

IfThen radar target to be measured belongs to V_iClass.

With regard to the classifying identification method of support vector machine, prior art literature " Melgani F, Bruzzone is can be found in L.Classification of hyperspectral remote sensing images with support vector machines.IEEE Trans.on Geoscince and Remote Sensing.2004,7(42):1778-1790”。

The present invention can be completely applied to be based on based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding The radar target recognition systems of computer programming self-operating, realize the radar target recognition of automatization.

Technical scheme is further described below by embodiment.

Embodiment：

The data image that the present embodiment is announced using MSTAR public databases, carrys out the comparative evaluation present invention based on Range Profile The SAR target identification methods of time-frequency figure non-negative sparse coding and the recognition effect of other Technology of Radar Target Identification.The present embodiment Ten class radar targets that MSTAR public databases publish be have chosen as the data of experimental data base.This ten classes radar mesh Mark is ground military vehicle or civilian vehicle, and external shape has similarity, and its radar target code name is respectively BMP2（Infantry Tank）、BRDM2（Amphibious armo(u)red scoutcar）、BTR60（Armoring carriage）、BTR70（Armored personnel carrier）、D7 （Agricultural bull-dozer）、T62（T-62 types main website tank）、T72（T-72 types main website tank）、ZIL131（Military trucks）、ZSU234 （Self propelled Antiaircraft Gun battlebus）And 2S1（Carriage motor howitzer battlebus）.The visible images of the ten classes radar target are respectively as corresponding in Fig. 4 Shown in the corresponding picture of code.In MSTAR public databases, this ten classes radar target that is stored with is in the different angles of pitch, difference Some radar target images that azimuth shoots, the present embodiment therefrom have chosen this ten classes radar target in 17 ° and 15 ° of pitching Tested in the part radar target image captured by multiple different orientations under angle, wherein, 17 ° of angles of pitch are shot The radar target image that 15 ° of angles of pitch shoot is made sample to be tested by training sample of the radar target image as experiment, to Carry out radar target recognition test.Choose in the experimental data base for obtaining in the present embodiment, the training sample of all kinds of radar targets Quantity and sample size to be tested are as shown in table 1.

Table 1

In the present embodiment, in order to compare, contrasted using three kinds of radar target identification methods, by the reality in terms of three Test and respectively the radar target recognition performance of these three radar target identification methods is evaluated and tested.Three kinds of radar target recognition sides Method is respectively：

Method 1., i.e. SAR target identification method of the present invention based on Range Profile time-frequency figure non-negative sparse coding, using this The non-negative sparse coding characteristic vector extracted to the Range Profile of SAR in bright method is known using support vector machine as identification feature Other algorithm carries out Classification and Identification to SAR targets.1. shorthand method is NNSC+SVM.

2., using the Range Profile to SAR method carries out Non-negative Matrix Factorization（non-negative matrix Factorization, is abbreviated as NMF）The nonnegative matrix characteristic vector extracted afterwards is used as identification feature（Referring to document " Du J X, Zhai C M,Ye Y Q.Face aging simulation and recognition based on NMF algorithm with sparseness constraints[J].Neurocomputing,2012,41,502-516”）, using supporting vector Machine recognizer carries out Classification and Identification to SAR targets.2. shorthand method is NMF+SVM.

3., using the Range Profile to SAR method carries out principal component analysiss（Principal Component Analysis, It is abbreviated as PCA）The principal component analysiss feature extracted afterwards as identification feature, using support vector machine recognizer to SAR targets Carry out Classification and Identification.3. shorthand method is PCA+SVM.

It should be noted that method is 1.（That is the inventive method）And do not need the bearing data of every width SAR image, i.e., not Prior estimation orientation angle is needed, and orientation angular estimation is very heavy for the classics SAR Target Recognition Algorithms such as template matching Want.

Experiment one：

First, by testing one come the above-mentioned three kinds of radar target identification methods of comparison to 10 classification target in experimental data base Recognition performance.As it was previously stated, before method NNSC feature extractions 1., needs carry out study according to training sample and obtain non- Negative dictionary matrix.In order to analyze impact of the non-negative dictionary matrix dimension to recognition result in NNSC, non-negative dictionary square is respectively provided with Battle array dimension is 10,20,30,40,50,60,70,80,90,100,120,150 and 200.Fig. 5 shows the non-negative that dimension is 100 Dictionary matrix.

Fig. 6 gives three kinds of radar target identification methods and 10 classification target recognition effects is become with non-negative dictionary matrix dimension The curve of change（In Fig. 6, abscissa is non-negative dictionary matrix dimension, and vertical coordinate is correct recognition rata）.From fig. 6, it can be seen that this The recognition accuracy of three kinds of methods improves with the increase of dimension.1. method is reached when non-negative dictionary matrix dimension is 80 To best discrimination, 2. method is also when non-negative dictionary matrix dimension is 80 to reach best discrimination；For side 3., best discrimination is when base dimension is 200 to method.However, for method 1., its discrimination is improved faster, Higher recognition accuracy can just be obtained under relatively low non-negative dictionary matrix dimension.When embodying the present invention based on Range Profile The recognition performance of the SAR target identification methods of frequency figure non-negative sparse coding is superior to other methods.

SAR target identification method of the present invention based on Range Profile time-frequency figure non-negative sparse coding, why than NMF+SVM side Method, PCA+SVM methods correct recognition rata it is higher, be that target office in time-frequency figure can effectively be extracted due to non-negative sparse coding The complicated scattering signatures in portion, and NMF and PCA methods are, from global angle extraction feature, to include excessive garbage.

Here, the SAR target identification methods and document also by the present invention based on Range Profile time-frequency figure non-negative sparse coding “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):The radar target identification method proposed in 1230-1242 " is identified Performance comparision.In document “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):In 1230-1242 ", Hidden Markov grader is combined based on Range Profile scattering center feature Method can reach optimal average correct recognition rata 92.16%, and the best identified rate of the inventive method is 98.78%, and this also illustrates Superiority of the present invention based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding.

Experiment two：

By experiment two, three kinds of radar recognition methodss are respectively adopted to 10 class radar mesh in the present embodiment experimental data base Mark carries out dubious recognition, compares the recognition performance of three kinds of recognition methodss.Dubious recognition, i.e., using every kind of radar identification side Method, based on the respective training sample of ten class radar targets in experimental data base, carries out thunder to the sample to be tested of ten class respectively Up to target recognition.

Table 2 gives method 1. when non-negative dictionary matrix dimension is 80 to 10 class thunders in the present embodiment experimental data base The identity confusion matrix and corresponding discrimination of dubious recognition are carried out up to target：

Table 2：Ten classification target identity confusion matrixes（NNSC+SVM, dictionary dimension 80）And discrimination

Table 3 gives method 2. when non-negative dictionary matrix dimension is 80 to 10 class thunders in the present embodiment experimental data base The identity confusion matrix and corresponding discrimination of dubious recognition are carried out up to target：

Table 3：Ten classification target identity confusion matrixes（NMF+SVM, dictionary dimension are 80）And discrimination

Table 4 gives method 3. when non-negative dictionary matrix dimension is 200 to 10 classes in the present embodiment experimental data base Radar target carries out the identity confusion matrix and corresponding discrimination of dubious recognition：

Table 4：Ten classification target identity confusion matrixes（PCA+SVM, dictionary dimension are 200）And discrimination

Table 2 to 4 shows that 1. method can reach the average recognition rate of highest 98.78%, and 2. method can reach highest 92.93% average recognition rate, and 3. method can only achieve the average recognition rate of highest 82.86%.It can be seen that in same instruction Practice with test set, recognition performance of the present invention based on the SAR target identification methods of Range Profile time-frequency figure non-negative sparse coding is excellent More in additive method.This also illustrates set forth herein method effectiveness and high-performance.

Experiment three：

By experiment three, recognition performance of three kinds of radar target identification methods under random noise environment is compared. In view of it there may be random noise in the case of, we increase Gaussian noise to each training and test data to assess three Plant the robustness of radar target identification method.

What Fig. 7 was represented is three kinds of radar target identification methods with signal to noise ratio（SNR）From -5db to 20db, range is being just The situation of change of ratio is recognized really.By Fig. 6, the recognition methodss of NNSC+SVM energy when SNR range 25-40db can be obtained Reach 90% accuracy rate, although the performance decline when SNR is less than 15db, but the recognition methodss of NNSC+SVM Discrimination in all SNR ranges under equal noise conditions is superior than other methods, when showing the present invention based on Range Profile The identification robustness of the SAR target identification methods of frequency figure non-negative sparse coding is still substantially superior to NMF+SVM and PCA+SVM and knows Other method, the recognition performance for possessing optimum.

By the comparison of above-mentioned recognition performance, it can be deduced that such as draw a conclusion.For the SAR object procedures based on Range Profile In, NNSC and NMF belongs to the Feature Extraction Technology based on part, the Feature Extraction Technology based on global change more this than PCA With more preferable correct recognition rata.Because the feature of target distance image time-frequency plane has significant openness, localization spy Point；It is better than based on global feature extraction effect based on the feature extraction effect of part.Know in the radar target based on part In other method, from experimental result it can be seen that NNSC is better than the effect of NMF.It should be noted that NNSC methods can be in relative base Higher discrimination is obtained in the case of dimension is less, and it is most significant mesh that its reason is the local feature that NNSC methods are extracted Mark substitutive characteristics, therefore dimension that need not be very high just can effectively differentiate target type, this point is also superior to NMF methods.

Finally illustrate, above example is only unrestricted to illustrate technical scheme, although with reference to compared with Good embodiment has been described in detail to the present invention, it will be understood by those within the art that, can be to the skill of the present invention Art scheme is modified or equivalent, and without deviating from the objective and scope of technical solution of the present invention, which all should be covered at this In the middle of the right of invention.

Claims (3)

1. SAR target identification methods based on Range Profile time-frequency figure non-negative sparse coding, it is characterised in that comprise the steps：

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

B adaptive Gaussian representation method is adopted), exploded representation is iterated to the Range Profile of SAR using Gaussian bases, is entered And it is calculated the time-frequency matrix of the Range Profile of SAR；Step B is specially：

Using adaptive Gaussian representation method, exploded representation is iterated to the Range Profile of SAR using Gaussian bases：

And

Wherein, t express times；The Range Profile of the SAR of r (t) express time ts；g_iT () represents to adjust the distance and carry out i-th as r (t) The Gaussian bases of secondary Breaking Recurrently, C_iExpression is adjusted the distance as r (t) carries out the expansion coefficient of ith iteration decomposition, i ∈ 1, 2,…,i_max, i_maxExpression is iterated the iteration total degree of exploded representation, and iteration total degree i_maxSo that the weight of Breaking Recurrently Structure errorMeetε is default reconstructed error threshold value, and 10^-3≤ε≤10^-5；And have relational expression：

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

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

t_i,f_i,α_iExpression to be adjusted the distance and carry out the resolution parameter in the Gaussian bases of ith iteration decomposition as r (t)；r_iT () represents Adjust the distance the iteration discrepance before ith iteration decomposition is carried out as r (t), and Represent Gaussian bases g_iThe conjugation of (t)；| | represent the operator that takes absolute value；Represent adjustment resolution parameter t_i,f_i,α_iSo that | |²Maximum maximum operator；

Above-mentioned relation formula is solved using Fourier Transform Algorithm, the resolution parameter t of ith iteration decomposition is obtained_i,f_i,α_iAnd expansion Coefficient C_i, and then time-frequency matrix Θ (t, f) of Range Profile r (t) is obtained by following formula：

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

Wherein, t express times, f represent frequency；

C) for the known radar target that multiclass is different, the SAR image of multiple known radar targets is gathered respectively as training sample This, and process the time-frequency matrix for obtaining each training sample in each classification according to step A～B respectively；Again by each training sample All non-negative pixel arrangements in this time-frequency matrix form a column vector, so as to the time-frequency matrix structure by each training sample Into one training sample matrix Z of column vector arrangement form；Using non-negative sparse coding learning algorithm by training sample matrix Z point Solve the product for non-negative dictionary matrix D and non-negative sparse coefficient matrix H：

Z=DH；

D) the non-negative dictionary matrix obtained using training sample matrix decomposition, is obtained according to the time-frequency matrix calculus of each training sample To the respective non-negative sparse coding characteristic vector of each training sample, so as to the non-negative sparse for obtaining all kinds of known radar targets is compiled Code total characteristic vector；Step D is specially：

For the time-frequency matrix of j-th training sample of pth class known radar targetObtained using training sample matrix decomposition Non-negative dictionary matrix D, calculate its corresponding non-negative sparse coding characteristic vector

$v_i^p={\left(D^{T}D\right)}^{-1}{D}^T{z}_j^p;$

(·)^-1Representing matrix inversion operation is accorded with；So as to the corresponding non-negative of all training samples by pth class known radar target is dilute Thin coding characteristic vector constitutes the non-negative sparse coding total characteristic vector V of pth class known radar target_p：

$V_p=\[v_1^p,v_2^p,...,v_j^p,...,v_J^p\];$

Wherein, J represents the training sample sum of pth class known radar target；Thus, all kinds of known radar targets are respectively obtained Non-negative sparse coding total characteristic vector；

E radar target to be measured is directed to), is gathered the SAR image of radar target to be measured, radar to be measured is obtained according to step A～B process The time-frequency matrix of target；The non-negative dictionary matrix obtained using training sample matrix decomposition, according to the time-frequency of radar target to be measured Matrix calculus obtain the non-negative sparse coding characteristic vector of radar target to be measured；

F) the non-negative sparse coding total characteristic vector using all kinds of known radar targets is known as identification benchmark using SVM Other algorithm, carries out Classification and Identification to the non-negative sparse coding characteristic vector of radar target to be measured, obtains the knowledge of radar target to be measured Other result.

2. SAR target identification methods according to claim 1 based on Range Profile time-frequency figure non-negative sparse coding, its feature exist In step C is specially：

C1) for the known radar target that multiclass is different, the SAR image of multiple known radar targets is gathered respectively as training sample This, and process the time-frequency matrix for obtaining each training sample in each classification according to step A～B respectively；

C2 all N number of non-negative pixel arrangement in the time-frequency matrix of each training sample is formed into a column vector), so as to by each The training sample matrix Z of one N × M of column vector arrangement form that the time-frequency matrix of individual training sample is constituted；M represents training sample Total number；

C3) training sample matrix Z is decomposed into non-negative dictionary matrix D and the K × M of N × K using non-negative sparse coding learning algorithm Non-negative sparse coefficient matrix H product：

Z=DH；

K represents the quantity of non-negative dictionary atom in non-negative sparse coding.

3. SAR target identification methods according to claim 1 based on Range Profile time-frequency figure non-negative sparse coding, its feature exist In step E is specially：

The SAR image of radar target to be measured is gathered, and the time-frequency matrix z of radar target to be measured is obtained according to step A～B process_x, so The non-negative dictionary matrix D for being obtained using training sample matrix decomposition afterwards, calculates the corresponding non-negative sparse coding of radar target to be measured Characteristic vector v_x：

v_x=(D^TD)^-1D^Tz_x；

(·)^-1Representing matrix inversion operation is accorded with.