CN103824093A - SAR (Synthetic Aperture Radar) image target characteristic extraction and identification method based on KFDA (Kernel Fisher Discriminant Analysis) and SVM (Support Vector Machine) - Google Patents

Abstract

The invention provides an SAR (Synthetic Aperture Radar) image target characteristic extraction and identification method based on KFDA (Kernel Fisher Discriminant Analysis) and an SVM (Support Vector Machine). The method comprises the following steps: performing amplitude data normalization processing on a training target sample of a known type and a testing target sample of an unknown type; performing characteristic extraction on the normalized training target sample of the known type and the testing target sample of the unknown type respectively by using a KFDA criterion; training an SVM classifier by using training target sample characteristics of known types extracted according to the KFDA criterion to generate an optimal classification face; identifying the characteristics of the testing target sample of the unknown type extracted according to the KFDA criterion through the optimal classification face. By adopting the method, the requirement on a preprocessing process is lowered, the target-aspect sensitivity of an SAR image is avoided, the dimensions of sample characteristics are compressed, and high target identification rate is obtained. The method has high popularity.

Description

A kind of feature extraction of SAR image object and recognition methods based on KFDA and SVM

Technical field

The invention belongs to SAR Image Processing and Pattern Recognition field, relate to a kind of based on KFDA(Kernel Fisher Discriminant Analysis) and SVM(Support Vector Machine) the feature extraction of SAR image object and recognition methods.

Background technology

Synthetic-aperture radar (Synthetic Aperture Radar, SAR) be a kind of active sensor that utilizes microwave perception, can carry out round-the-clock, round-the-clock scouting to interesting target or region, there is various visual angles, the acquisition capability of many angles of depression data and the penetration capacity to some atural objects.So-called radar target recognition, at radar, target is detected exactly and the basis of locating on, according to the radar echo signal of target and environment, extract target signature, the judgement of attribute, classification or the model of realize target.Along with the continuous maturation of SAR imaging technique, the target identification based on SAR image has more and more important meaning.

In target identification process based on SAR image, most important two steps are feature extraction and identification.For SAR image, due to its special imaging mode, make it unlike optical imagery the global shape of target can more intactly be described, distribute but show as sparse scattering center, and comparatively responsive to the orientation of imaging.Therefore, effectively extract target signature and seem particularly important.After having obtained the clarification of objective of SAR image, ensuing main task is identified unknown object exactly.

In the target's feature-extraction method of SAR image, the most frequently used is the methods such as the principal component analysis (KPCA) of principal component analysis (PCA), kernel function, wherein the shortcoming of principal component method is to extract the nonlinear characteristic existing in image, and the shortcoming of the principal component method of kernel function is extracted feature, not have the dimension of good class discriminating power and feature higher; In the target identification method of SAR image, the most frequently used is maximal correlation sorter and nearest neighbor classifier etc., wherein the shortcoming of maximal correlation sorter is in the time that sample dimension is higher, and algorithm complex is also higher, and the shortcoming of nearest neighbor classifier is the optimal classification Mian Bushi global optimum choosing.

Summary of the invention

The technical problem to be solved in the present invention is: a kind of feature extraction of SAR image object and recognition methods based on KFDA and SVM is provided, and the method utilizes KFDA criterion to carry out target's feature-extraction, then by the identification of svm classifier device realize target.The present invention is by combining KPCA criterion with svm classifier device, the present invention can very well complete target's feature-extraction and the identification of SAR image, reduce the requirement to preprocessing process, overcome the azimuthal sensitivity of SAR image, compress the dimension of sample characteristics, and obtain higher object recognition rate, there is good generalization.

The technical solution adopted for the present invention to solve the technical problems is:

The feature extraction of SAR image object and a recognition methods based on KFDA and SVM, comprise following step:

The training objective sample of step (1) to known class and the test target sample of unknown classification carry out amplitude data normalized;

The training objective sample of known class and the test target sample data of unknown classification after step (2) utilizes KFDA criterion to normalization are carried out feature extraction;

The training objective sample characteristics of known class that step (3) utilizes KFDA criterion to extract, to SVM(Support Vector Machine) sorter trains, and produces optimal classification face;

Step (4) is by optimal classification face, and the feature of the test target sample of the unknown classification that KFDA criterion is extracted is identified.

Further, the process that in described step (1), the test target sample of the training objective sample to known class and unknown classification carries out amplitude data normalized is specially:

Normalization formula is:

$x_{\mathrm{Normalized}}=\frac{x}{{\left|\right|x\left|\right|}_2}$

Wherein, the vector representation (being arranged in vector form by row by image array) of the training objective sample that x is any known class or the test target sample of unknown classification, x _normalizedfor the vector representation after the amplitude data normalization of the training objective sample of corresponding known class or the test target sample of unknown classification.

Further, the process that the training objective sample of known class after utilizing KFDA criterion to normalization in described step (2) and the test target sample data of unknown classification are carried out feature extraction is specially: first ask Scatter Matrix K in class _wwith between class scatter matrix K _b, then ask

nonzero eigenvalue characteristic of correspondence vector, finally ask the feature of the training objective sample of the known class under KFDA criterion and the test target sample of unknown classification; Wherein Scatter Matrix K in class _wfor:

$K_w=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{K}_i(I-\frac{1}{N_i})K_i^T$

Wherein, the number of the training objective sample that N is known class, the classification number of the training objective sample that c is known class,

for N × N _imatrix, x _p(p=1,2 ..., N) and be the data after the training objective sample normalization of p known class,

be the data after the normalization of j training objective sample in i class, N _ibe the sample number of the training objective sample of i class known class, k ₁() represents kernel function, and I is N _i× N _iunit matrix, for element is

n _i× N _isquare formation; If K _wmatrix is a singular matrix, makes K _w≈ K _w+ κ I is to solve K _wsingularity, I is and K _wthe unit matrix of same order, κ be one very little and be greater than zero disturbance constant, common desirable κ≤10 ^-2;

Between class scatter matrix K _bfor:

$K_b=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{N}_i(G_i-G_0){(G_i-G_0)}^T$

Wherein, $G_i={(\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_1,x_j^i),\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_2,x_j^{\text{i}}),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_N,x_j^i))}^T,x_j^i(j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N_i)$ Be the data after the normalization of j training objective sample in i class, ${\mathrm{<figure-callout\; id="0"\; label="G"\; filenames="HDA0000479324470000021.png"\; state="\{\{state\}\}">G</figure-callout>}}_0={(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_1,x_j),\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_2,x_j),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_N,x_j))}^T.$

Ask again

nonzero eigenvalue characteristic of correspondence vector α, that is:

$\mathrm{\&lambda;\&alpha;}=K_w^{-1}{K}_b\mathrm{\&alpha;}$

Wherein, λ representation feature value; Training objective sample to the known class after any normalization or the test target sample of unknown classification, the final training objective sample of known class or the test target sample characteristics of unknown classification extracting of KFDA criterion is a c-1 dimensional vector, can be expressed as z=[z ₁, z ₂..., z _c-1] ^t, every one-dimensional element can be expressed as:

$z_t=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_j^t{k}_1(x_j,x)$

Wherein, t=1,2 ..., c-1,

represent

j element of t nonzero eigenvalue institute character pair vector.

Further, utilize the training objective sample characteristics of the known class of KFDA criterion extraction in described step (3), svm classifier device is trained, the process that produces optimal classification face is specially:

Utilize Lagrange multiplier method, maximize functional:

$Q\left(a\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{a}_j{y}_i{y}_j{k}_2(z_i,z_j)$

$s.t.\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{a}_i=0$

a _i≥0,i＝1,…,n

Wherein, y _i∈+1 ,-1}, respectively the training objective sample of two corresponding different class known class; z _ithe feature of training objective of i the known class of extracting for KFDA criterion; k ₂() represents kernel function, with above-mentioned k ₁() is completely independent; a _ifor the corresponding Lagrange multiplier of the training objective sample with i known class to be asked.

Further, in described step (4), by optimal classification face, the process that the feature of the test target sample of the unknown classification that KFDA criterion is extracted is identified is specially:

The function representation of optimal classification face is:

$f\left(x\right)=\mathrm{sgn}\{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{y}_i{k}_2(z_i,z)+b^{*}\}$

Wherein, a _ifor the corresponding Lagrange multiplier of the training objective sample with i known class to be asked; y _i∈+1 ,-1}, respectively the training objective sample of two corresponding different class known class; k ₂() represents kernel function; z _ithe feature of training objective of i the known class of extracting for KFDA criterion; Z is the feature of the test target sample of the unknown classification of KFDA criterion extraction;

{+1 ,-1}, has determined the affiliated classification of the test target sample of current unknown classification to f (x) ∈.

Principle of the present invention is: core Fisher criterion is applied to feature extraction, in higher dimensional space, extracts the nonlinear characteristic of SAR image by kernel method, in the hope of obtaining better class discriminating power; What SVM found is an optimum lineoid that meets classificating requirement, is applicable to small sample, the problem such as non-linear; So by the two superiority combination, can realize well target's feature-extraction and the identification of SAR image.

The present invention's advantage is compared with prior art:

1. for target's feature-extraction part, compare to PCA criterion, the present invention can obtain the nonlinear characteristic in image;

2. for target's feature-extraction part, compare to KPCA criterion, the present invention can obtain lower intrinsic dimensionality, and has better robustness;

3. for target identification division, compare to maximal correlation sorter, the present invention has solved problem of dimension dexterously, and algorithm complex and sample dimension are irrelevant;

4. for target identification division, compare to nearest neighbor classifier, what the present invention obtained is globally optimal solution;

5. by KPCA criterion is combined with svm classifier device, the present invention can very well complete target's feature-extraction and the identification of SAR image, reduce the requirement to preprocessing process, overcome the azimuthal sensitivity of SAR image, compress the dimension of sample characteristics, and obtain higher object recognition rate, there is good generalization.

Accompanying drawing explanation

Fig. 1 is target's feature-extraction of the present invention and identification process figure;

Fig. 2 is the process of example being carried out to target's feature-extraction and identification.

Wherein:

(a) a, b while being 17 ° for the angle of pitch, c tri-class tanks are as the two dimensional character distribution plan of the training objective sample of known class;

(b) the two dimensional character distribution plan when c class tank when the angle of pitch is 15 ° is identified;

(c) the sex change target c#(model of the c class tank when the angle of pitch is 15 ° is identical, configures different targets) two dimensional character distribution plan while identifying.

Embodiment

Introduce in detail the present invention below in conjunction with the drawings and the specific embodiments.

As shown in Figure 1, the concrete implementation step of SAR image object feature extraction based on KFDA and SVM of the present invention and recognition methods is as follows:

Step (1), training objective sample to known class and the test target sample of unknown classification carry out amplitude data normalized, and normalization formula is:

$x_{\mathrm{Normalized}}=\frac{x}{{\left|\right|x\left|\right|}_2}$

Wherein, the vector representation (being arranged in vector form by row by image array) of the training objective sample that x is any known class or the test target sample of unknown classification, x _normalizedfor the vector representation after the amplitude data normalization of the training objective sample of corresponding known class or the test target sample of unknown classification.

Step (2), the training objective sample of known class and the test target sample data x of unknown classification after utilizing KFDA criterion to normalization _normalizedcarry out feature extraction, first ask Scatter Matrix K in class _wwith between class scatter matrix K _b, then ask nonzero eigenvalue characteristic of correspondence vector, finally ask the feature of the training objective sample of the known class under KFDA criterion and the test target sample of unknown classification; Wherein Scatter Matrix K in class _wfor:

$K_w=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{K}_i(I-\frac{1}{N_i})K_i^T$

Wherein, the number of the training objective sample that N is known class, the classification number of the training objective sample that c is known class,

for N × N _imatrix, x _p(p=1,2 ..., N) and be the data after the training objective sample normalization of p known class,

be the data after the normalization of j training objective sample in i class, N _ibe the sample number of the training objective sample of i class known class, k ₁() represents kernel function, and I is N _i× N _iunit matrix,

for element is

n _i× N _isquare formation; If K _wmatrix is a singular matrix, makes K _w≈ K _w+ κ I is to solve K _wsingularity, I is and K _wthe unit matrix of same order, κ be one very little and be greater than zero disturbance constant, common desirable κ≤10 ^-2;

Between class scatter matrix K _bfor:

$K_b=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{N}_i(G_i-G_0){(G_i-G_0)}^T$

Wherein, $G_i={(\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_1,x_j^i),\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_2,x_j^{\text{i}}),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_N,x_j^i))}^T,x_j^i(j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N_i)$ Be the data after the normalization of j training objective sample in i class, ${\mathrm{<figure-callout\; id="0"\; label="G"\; filenames="HDA0000479324470000021.png"\; state="\{\{state\}\}">G</figure-callout>}}_0={(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_1,x_j),\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_2,x_j),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_N,x_j))}^T.$

Ask again

nonzero eigenvalue characteristic of correspondence vector α, that is:

$\mathrm{\&lambda;\&alpha;}=K_w^{-1}{K}_b\mathrm{\&alpha;}$

Wherein, λ representation feature value; Training objective sample to the known class after any normalization or the test target sample of unknown classification, the final training objective sample of known class or the test target sample characteristics of unknown classification extracting of KFDA criterion is a c-1 dimensional vector, can be expressed as z=[z ₁, z ₂..., z _c-1] ^t, every one-dimensional element can be expressed as:

$z_t=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_j^t{k}_1(x_j,x)$

Wherein, t=1,2 ..., c-1,

represent

j element of t nonzero eigenvalue institute character pair vector.

Step (3), the training objective sample characteristics of known class that utilizes KFDA criterion to extract, train svm classifier device, produces optimal classification face, can utilize Lagrange multiplier method, maximizes functional:

$Q\left(a\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{a}_j{y}_i{y}_j{k}_2(z_i,z_j)$

$s.t.\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{a}_i=0$

a _i≥0,i＝1,…,n

Wherein, y _i∈+1 ,-1}, respectively the training objective sample of two corresponding different class known class; z _ithe feature of training objective of i the known class of extracting for KFDA criterion; k ₂() represents kernel function, with above-mentioned k ₁() is completely independent; a _ifor the corresponding Lagrange multiplier of the training objective sample with i known class to be asked.

Step (4), by optimal classification face, the feature of the test target sample of the unknown classification that KFDA criterion is extracted is identified, wherein, the function representation of optimal classification face is:

$f\left(x\right)=\mathrm{sgn}\{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{y}_i{k}_2(z_i,z)+b^{*}\}$

Wherein, a _ifor the corresponding Lagrange multiplier of the training objective sample with i known class to be asked; y _i∈+1 ,-1}, respectively the training objective sample of two corresponding different class known class; k ₂() represents kernel function; z _ithe feature of training objective of i the known class of extracting for KFDA criterion; Z is the feature of the test target sample of the unknown classification of KFDA criterion extraction;

{+1 ,-1}, has determined the affiliated classification of the test target sample of current unknown classification to f (x) ∈.

201 a, b while being 17 ° for the angle of pitch in Fig. 2, c tri-class tanks are as the two dimensional character distribution plan of the training objective sample of known class; 202 two dimensional character distribution plans while identifying for the c class tank when the angle of pitch is 15 °; Two dimensional character distribution plan when sex change target c# class (model is identical, the configures different targets) tank of 203 c class tanks when the angle of pitch is 15 ° is identified.Can find out, in Figure 20 2, as same configuration target, the two dimensional character of the test target sample c class tank of unknown classification is gathered in around the two dimensional character of training objective sample c class tank of known class well; In Figure 20 3, as Morph Target, the two dimensional character of the test target sample c# class tank of unknown classification is also gathered in around the two dimensional character of training objective sample c class tank of known class well.Prove that the method has superior target recognition capability, realized the feature extraction of SAR image object and identification.

The content not being described in detail in instructions of the present invention belongs to the known prior art of professional and technical personnel in the field.

Although disclose for the purpose of illustration most preferred embodiment of the present invention and accompanying drawing, it will be appreciated by those skilled in the art that: without departing from the spirit and scope of the invention and the appended claims, various replacements, variation and modification are all possible.Therefore the technical scheme that, the present invention protects should not be limited to most preferred embodiment and the disclosed content of accompanying drawing.

Claims (5)

1. the feature extraction of SAR image object and the recognition methods based on KFDA and SVM, is characterized in that, comprises following step:

The training objective sample of step (1) to known class and the test target sample of unknown classification carry out amplitude data normalized;

The training objective sample of known class and the test target sample data of unknown classification after step (2) utilizes KFDA criterion to normalization are carried out feature extraction;

Step (3) is utilized the training objective sample characteristics of the known class of KFDA criterion extraction, and svm classifier device is trained, and produces optimal classification face;

Step (4) is by optimal classification face, and the feature of the test target sample of the unknown classification that KFDA criterion is extracted is identified.

2. the feature extraction of SAR image object and the recognition methods based on KFDA and SVM according to claim 1, is characterized in that: the process that the training objective sample to known class in described step (1) and the test target sample of unknown classification carry out amplitude data normalized is specially:

Normalization formula is:

$x_{\mathrm{Normalized}}=\frac{x}{{\left|\right|x\left|\right|}_2}$

Wherein, the vector representation (being arranged in vector form by row by image array) of the training objective sample that x is any known class or the test target sample of unknown classification, x _normalizedfor the vector representation after the amplitude data normalization of the training objective sample of corresponding known class or the test target sample of unknown classification.

3. the feature extraction of SAR image object and the recognition methods based on KFDA and SVM according to claim 1, is characterized in that: the process that the training objective sample of known class after utilizing KFDA criterion to normalization in described step (2) and the test target sample data of unknown classification are carried out feature extraction is specially: first ask Scatter Matrix K in class _wwith between class scatter matrix K _b, then ask

nonzero eigenvalue characteristic of correspondence vector, finally ask the feature of the training objective sample of the known class under KFDA criterion and the test target sample of unknown classification; Wherein Scatter Matrix K in class _wfor:

$K_w=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{K}_i(I-\frac{1}{N_i})K_i^T$

Wherein, the number of the training objective sample that N is known class, the classification number of the training objective sample that c is known class,

for N × N _imatrix, x _p(p=1,2 ..., N) and be the data after the training objective sample normalization of p known class,

be the data after the normalization of j training objective sample in i class, N _ibe the sample number of the training objective sample of i class known class, k ₁() represents kernel function, and I is N _i× N _iunit matrix,

for element is

n _i× N _isquare formation; If K _wmatrix is a singular matrix, makes K _w≈ K _w+ κ I is to solve K _wsingularity, I is and K _wthe unit matrix of same order, κ be one very little and be greater than zero disturbance constant, common desirable κ≤10 ^-2;

Between class scatter matrix K _bfor:

$K_b=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{N}_i(G_i-G_0){(G_i-G_0)}^T$

Wherein, $G_i={(\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_1,x_j^i),\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_2,x_j^{\text{i}}),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}{k}_1(x_N,x_j^i))}^T,x_j^i(j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,N_i)$ Be the data after the normalization of j training objective sample in i class, $G_0={(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_1,x_j),\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_2,x_j),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{k}_1(x_N,x_j))}^T;$

Ask again

nonzero eigenvalue characteristic of correspondence vector α, that is:

$\mathrm{\&lambda;\&alpha;}=K_w^{-1}{K}_b\mathrm{\&alpha;}$

Wherein, λ representation feature value; Training objective sample to the known class after any normalization or the test target sample of unknown classification, the final training objective sample of known class or the test target sample characteristics of unknown classification extracting of KFDA criterion is a c-1 dimensional vector, can be expressed as z=[z ₁, z ₂..., z _c-1] ^t, every one-dimensional element can be expressed as:

$z_t=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_j^t{k}_1(x_j,x)$

Wherein, t=1,2 ..., c-1,

represent

j element of t nonzero eigenvalue institute character pair vector.

4. the feature extraction of SAR image object and the recognition methods based on KFDA and SVM according to claim 3, it is characterized in that: the training objective sample characteristics that utilizes the known class of KFDA criterion extraction in described step (3), svm classifier device is trained, and the process that produces optimal classification face is specially:

Utilize Lagrange multiplier method, maximize functional:

$Q\left(a\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{a}_j{y}_i{y}_j{k}_2(z_i,z_j)$

$s.t.\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{a}_i=0$

a _i≥0,i＝1,…,n

Wherein, y _i∈+1 ,-1}, respectively the training objective sample of two corresponding different class known class; z _ithe feature of training objective of i the known class of extracting for KFDA criterion; k ₂() represents kernel function, with above-mentioned k ₁() is completely independent; a _ifor the corresponding Lagrange multiplier of the training objective sample with i known class to be asked.

5. the feature extraction of SAR image object and the recognition methods based on KFDA and SVM according to claim 1, it is characterized in that: in described step (4), by optimal classification face, the process that the feature of the test target sample of the unknown classification that KFDA criterion is extracted is identified is specially:

The function representation of optimal classification face is:

$f\left(x\right)=\mathrm{sgn}\{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{y}_i{k}_2(z_i,z)+b^{*}\}$

Wherein, a _ifor the corresponding Lagrange multiplier of the training objective sample with i known class to be asked; y _i∈+1 ,-1}, respectively the training objective sample of two corresponding different class known class; k ₂() represents kernel function; z _ithe feature of training objective of i the known class of extracting for KFDA criterion; Z is the feature of the test target sample of the unknown classification of KFDA criterion extraction;

{+1 ,-1}, has determined the affiliated classification of the test target sample of current unknown classification to f (x) ∈.

Images