CN103824093A

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

Info

Publication number: CN103824093A
Application number: CN201410103639.2A
Authority: CN
Inventors: 高飞; 梅净缘; 孙进平; 王俊; 吕文超
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2014-03-19
Filing date: 2014-03-19
Publication date: 2014-05-28
Anticipated expiration: 2034-03-19
Also published as: CN103824093B

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

SAR image target feature extraction and recognition method based on KFDA and SVM

Technical Field

The invention belongs to the field of SAR image processing and pattern recognition, and relates to an SAR image target feature extraction and recognition method based on KFDA (Kernel Fisher cognitive analysis) and SVM (support Vector machine).

Background

Synthetic Aperture Radar (SAR) is an active sensor using microwave sensing, can perform all-day and all-weather reconnaissance on an interested target or area, and has the capability of acquiring multi-view and multi-depression data and the capability of penetrating ground objects. The radar target identification is to extract target features according to radar echo signals of a target and an environment on the basis of detection and positioning of the target by a radar, so as to judge the attribute, the category or the model of the target. With the continuous maturation of the SAR imaging technology, the target identification based on the SAR image has more and more important significance.

In the SAR image-based target identification process, the most important two steps are feature extraction and identification. Due to the special imaging mode of the SAR image, the SAR image can not describe the overall shape of a target completely like a general optical image, but shows sparse scattering center distribution and is sensitive to the imaging direction. Therefore, it is important to effectively extract the target features. After the features of the target of the SAR image are obtained, the next main task is to identify the unknown target.

In the method for extracting target features of the SAR image, methods such as Principal Component Analysis (PCA), kernel function principal component analysis (KPCA) and the like are most commonly used, wherein the principal component analysis method has the disadvantage that non-linear features existing in the image cannot be extracted, and the kernel function principal component analysis method has the disadvantage that the extracted features do not have good class discrimination capability and the feature dimension is high; in the target identification method of the SAR image, a maximum correlation classifier and a nearest neighbor classifier and the like are most commonly used, wherein the maximum correlation classifier has the disadvantage that the algorithm complexity is higher when the sample dimension is higher, and the nearest neighbor classifier has the disadvantage that the selected optimal classification surface is not globally optimal.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the method comprises the steps of extracting target features by using a KFDA criterion and realizing target recognition through an SVM classifier. By combining the KPCA criterion with the SVM classifier, the method can well complete the target feature extraction and identification of the SAR image, reduce the requirement on the preprocessing process, overcome the orientation sensitivity of the SAR image, compress the dimension of the sample feature, obtain higher target identification rate and have good popularization.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a method for extracting and identifying SAR image target features based on KFDA and SVM comprises the following steps:

step (1) carrying out amplitude data normalization processing on training target samples of known types and test target samples of unknown types;

performing feature extraction on the normalized training target sample of the known type and the test target sample data of the unknown type by using a KFDA (Kalman Filter/data acquisition) criterion;

training an SVM (support vector machine) classifier by using training target sample characteristics of known types extracted by a KFDA (KFDA) criterion to generate an optimal classification surface;

and (4) identifying the characteristics of the test target sample of unknown class extracted by the KFDA criterion through the optimal classification surface.

Further, the process of performing amplitude data normalization processing on the training target samples of the known type and the test target samples of the unknown type in the step (1) specifically includes:

the normalized formula is:

x_{Normalized} = \frac{x}{{| | x | |}_{2}}

where x is a vector representation of a training target sample of any known class or a test target sample of unknown class (i.e., the image matrix is arranged in columns in vector form), x_NormalizedAnd (3) vector representation after normalization of the amplitude data of the training target sample of the corresponding known class or the testing target sample of the unknown class.

Further, the process of performing feature extraction on the normalized training target samples of the known type and the test target sample data of the unknown type by using the KFDA criterion in the step (2) specifically comprises: first solving an intra-class divergence matrix K_wAnd the inter-class divergence matrix K_bThen ask for

Finally solving the characteristics of the training target sample of the known type and the test target sample of the unknown type under the KFDA criterion; wherein the intra-class divergence matrix K_wComprises the following steps:

wherein N is the number of training target samples of known classes, c is the number of classes of training target samples of known classes,

is NxN_iMatrix, x_p(p is 1,2, …, N) is the normalized data of the training target sample of the p-th known class,

normalized data for the jth training target sample in class i, N_iNumber of samples, k, of training target samples of known class i₁(-) represents a kernel function, I is N_i×N_iThe unit array is formed by a plurality of unit arrays,is composed of elements of

N of (A)_i×N_iA square matrix; if K_wIf the matrix is a singular matrix, let K_w≈K_w+ kappa I to solve K_wSingularity of (A) and (B)_wUnity matrix of the same order, κ is a perturbation constant small and greater than zero, and it is generally desirable that κ ≦ 10^-2；

Inter-class divergence matrix K_bComprises the following steps:

wherein,

normalized data for the jth training target sample in class i,

then ask for

The non-zero eigenvalue of (a) corresponds to the eigenvector α, i.e.:

wherein λ represents a characteristic value; for any normalized training target sample of a known class or test target sample of an unknown class, the characteristics of the training target sample of the known class or the test target sample of the unknown class finally extracted by the KFDA criterion are all a c-1-dimensional vector, and can be expressed as z ═ z [ -z [₁,z₂,…,z_c-1]^TEach dimension element may be represented as:

wherein t is 1,2, …, c-1,

to represent

The jth non-zero eigenvalue of (a) corresponds to the jth element of the eigenvector.

Further, the training of the SVM classifier is performed by using the training target sample features of known classes extracted by the KFDA criterion in step (3), and the process of generating the optimal classification surface specifically includes:

maximizing the functional by Lagrange multiplier method:

a_i≥0,i＝1,…,n

wherein, y_iThe method comprises the following steps that (1) a { +1, -1} is formed, and training target samples of two different types of known classes are respectively corresponding to the samples; z is a radical of_iFeatures of an ith known class of training target extracted for the KFDA criteria; k is a radical of₂(-) represents a kernel function, with k above₁(-) is completely independent; a is_iIs the Lagrange multiplier to be solved corresponding to the training target sample of the ith known class.

Further, the process of identifying the features of the test target samples of unknown types extracted by the KFDA criterion through the optimal classification surface in the step (4) specifically includes:

the function of the optimal classification surface is expressed as:

wherein, a_iThe Lagrange multiplier corresponding to the training target sample of the ith known class to be solved; y is_iThe method comprises the following steps that (1) a { +1, -1} is formed, and training target samples of two different types of known classes are respectively corresponding to the samples; k is a radical of₂(-) represents a kernel function; z is a radical of_iFor training targets of the ith known class extracted for the KFDA criterionCharacteristic; z is the characteristic of the test target sample of unknown class extracted by the KFDA criterion;

f (x) E { +1, -1}, i.e. the class of the test target sample of the current unknown class is determined.

The principle of the invention is as follows: applying a kernel Fisher discrimination criterion to feature extraction, namely extracting nonlinear features of the SAR image in a high-dimensional space by a kernel method to obtain better class discrimination capability; the SVM searches for an optimal hyperplane meeting the classification requirement, and is suitable for the problems of small samples, nonlinearity and the like; therefore, the advantages of the SAR image and the target feature extraction and identification can be well realized by combining the advantages of the SAR image and the target feature extraction.

Compared with the prior art, the invention has the advantages that:

1. for the target feature extraction part, compared with the PCA criterion, the method can obtain the nonlinear features in the image;

2. for the target feature extraction part, compared with a KPCA (kernel principal component analysis) criterion, the method can obtain lower feature dimension and has better robustness;

3. for the target identification part, compared with the maximum correlation classifier, the method skillfully solves the problem of dimension, and the algorithm complexity is irrelevant to the sample dimension;

4. for the target identification part, compared with the nearest neighbor classifier, the global optimal solution is obtained by the method;

5. by combining the KPCA criterion with the SVM classifier, the method can well complete the target feature extraction and recognition of the SAR image, reduce the requirement on the preprocessing process, overcome the orientation sensitivity of the SAR image, compress the dimension of the sample feature, obtain higher target recognition rate and have good popularization.

Drawings

FIG. 1 is a flow chart of target feature extraction and identification in accordance with the present invention;

FIG. 2 is a process of target feature extraction and identification for an example.

Wherein:

(a) the method comprises the following steps of (1) taking three types of tanks a, b and c with a pitch angle of 17 degrees as a two-dimensional characteristic distribution map of a training target sample of a known type;

(b) the two-dimensional characteristic distribution map is used for identifying the c-type tank with the pitch angle of 15 degrees;

(c) the two-dimensional characteristic distribution diagram is used for identifying the degeneration target c # (namely, targets with the same model and different configurations) of the c-type tank with the pitching angle of 15 degrees.

Detailed Description

The invention is described in detail below with reference to the figures and the detailed description.

As shown in fig. 1, the method for extracting and identifying the target features of the SAR image based on KFDA and SVM of the present invention comprises the following steps:

step (1), carrying out amplitude data normalization processing on training target samples of known types and test target samples of unknown types, wherein the normalization formula is as follows:

x_{Normalized} = \frac{x}{{| | x | |}_{2}}

where x is a training target of any known classVector representation of samples or unknown classes of test target samples (i.e. arranging the image matrix in vector form by columns), x_NormalizedAnd (3) vector representation after normalization of the amplitude data of the training target sample of the corresponding known class or the testing target sample of the unknown class.

Step (2), training target samples of known types and test target sample data x of unknown types are normalized by using a KFDA (Kalman Filter-Kalman Filter) criterion_NormalizedExtracting features, and calculating the divergence matrix K in class_wAnd the inter-class divergence matrix K_bThen ask forFinally solving the characteristics of the training target sample of the known type and the test target sample of the unknown type under the KFDA criterion; wherein the intra-class divergence matrix K_wComprises the following steps:

normalized data for the jth training target sample in class i, N_iNumber of samples, k, of training target samples of known class i₁(-) represents a kernel function, I is N_i×N_iThe unit array is formed by a plurality of unit arrays,

is composed of elements of

Inter-class divergence matrix K_bComprises the following steps:

wherein,

normalized data for the jth training target sample in class i,

then ask for

The non-zero eigenvalue of (a) corresponds to the eigenvector α, i.e.:

wherein t is 1,2, …, c-1,

to represent

And (3) training the SVM classifier by using the training target sample characteristics of known types extracted by a KFDA (Kalman Filter and data acquisition) criterion to generate an optimal classification surface, namely maximizing the functional by using a Lagrange multiplier method:

a_i≥0,i＝1,…,n

wherein, y_iThe method comprises the following steps that (1) a { +1, -1} is formed, and training target samples of two different types of known classes are respectively corresponding to the samples; z is a radical of_iFeatures of an ith known class of training target extracted for the KFDA criteria; k is a radical of₂(-) represents a kernel function, with k above₁(-) is completely independent; a is_iIs on demandLagrange multipliers corresponding to training target samples of the ith known class.

And (4) identifying the characteristics of the test target sample of unknown class extracted by the KFDA criterion through the optimal classification surface, wherein the function of the optimal classification surface is expressed as:

wherein, a_iThe Lagrange multiplier corresponding to the training target sample of the ith known class to be solved; y is_iThe method comprises the following steps that (1) a { +1, -1} is formed, and training target samples of two different types of known classes are respectively corresponding to the samples; k is a radical of₂(-) represents a kernel function; z is a radical of_iFeatures of an ith known class of training target extracted for the KFDA criteria; z is the characteristic of the test target sample of unknown class extracted by the KFDA criterion;

In fig. 2, 201 is a two-dimensional feature distribution diagram of three types a, b and c of tanks with a pitch angle of 17 ° as training target samples of known types; 202 is a two-dimensional characteristic distribution map for identifying a c-type tank with a pitch angle of 15 degrees; reference numeral 203 denotes a two-dimensional feature distribution map for identifying a tank of the degeneration target type c # of the tank of type c (i.e., targets having the same model and different configurations) with a pitch angle of 15 °. It can be seen that, in the diagram 202, as the same configuration target, the two-dimensional features of the class c tank of the test target sample of unknown class are well gathered around the two-dimensional features of the class c tank of the training target sample of known class; in fig. 203, as a deformation target, the two-dimensional features of the test target sample class c # tank of the unknown class are also well clustered around the two-dimensional features of the training target sample class c tank of the known class. The method is proved to have excellent target identification capability, and the SAR image target feature extraction and identification are realized.

Those skilled in the art will appreciate that the invention may be practiced without these specific details.

Although the preferred embodiments of the present invention and the accompanying drawings have been disclosed for illustrative purposes, those skilled in the art will appreciate that: various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the technical solution protected by the present invention should not be limited to the disclosure of the best embodiment and the accompanying drawings.

Claims

1. A SAR image target feature extraction and recognition method based on KFDA and SVM is characterized by comprising the following steps:

training the SVM classifier by using the training target sample characteristics of known classes extracted by a KFDA (Kalman Filter and data acquisition) criterion to generate an optimal classification surface;

2. The method for extracting and identifying the target features of the SAR image based on the KFDA and the SVM of claim 1, wherein: the process of performing amplitude data normalization processing on the training target samples of the known type and the test target samples of the unknown type in the step (1) specifically comprises the following steps:

the normalized formula is:

x_{Normalized} = \frac{x}{{| | x | |}_{2}}

3. The method for extracting and identifying the target features of the SAR image based on the KFDA and the SVM of claim 1, wherein: the process of performing feature extraction on the normalized training target sample of the known type and the test target sample data of the unknown type by using the KFDA criterion in the step (2) specifically comprises the following steps: first solving an intra-class divergence matrix K_wAnd the inter-class divergence matrix K_bThen ask for

is composed of elements of

Inter-class divergence matrix K_bComprises the following steps:

wherein,

normalized data for the jth training target sample in class i,

then ask for

The non-zero eigenvalue of (a) corresponds to the eigenvector α, i.e.:

wherein t is 1,2, …, c-1,

to represent

4. The method for extracting and identifying the target features of the SAR image based on the KFDA and the SVM of claim 3, wherein: in the step (3), training the SVM classifier by using the training target sample characteristics of known types extracted by the KFDA criterion, wherein the process of generating the optimal classification surface specifically comprises the following steps:

maximizing the functional by Lagrange multiplier method:

a_i≥0,i＝1,…,n

5. The method for extracting and identifying the target features of the SAR image based on the KFDA and the SVM of claim 1, wherein: the process of identifying the features of the test target sample of unknown type extracted by the KFDA criterion through the optimal classification surface in the step (4) is specifically as follows:

the function of the optimal classification surface is expressed as: