CN102122352B - Characteristic value distribution statistical property-based polarized SAR image classification method - Google Patents

Abstract

The invention discloses a characteristic value Gaussian statistical property-based polarized synthetic aperture radar (SAR) image classification method, which mainly solves the problems that the prior art is insufficient on cognition of characteristic distribution properties and the category judgment limit needs man-made determination. The method comprises the following steps of: 1) performing characteristic value decomposition on all pixel points of polarized SAR images to be classified; 2) selecting different homogeneous regions as the most basic category representative regions, and extracting characteristic values for representing the homogeneous regions; 3) estimating Gaussian hybrid model parameters of the characteristic values lambda 1, lambda 2 and lambda 3 of the homogeneous regions by adopting an expectation-maximization (EM) algorithm respectively, and solving a probability density distribution function of the characteristic values; 4) solving a joint probability distribution function of three characteristic values of each homogeneous region; and 5) performing Bayesian classification on the pixel points in the homogeneous regions, and outputting the classification results. The method has the advantage of remarkable classification effect on the polarized SAR images, and can be used for target detection and target identification of the polarized SAR images.

Description

Polarization SAR image classification method based on the eigenvalue distribution statistical property

Technical field

The invention belongs to technical field of image processing, relate to the polarization SAR image classification, can be used for Radar Targets'Detection and Target Recognition.

Background technology

Synthetic-aperture radar SAR utilizes the synthetic aperture principle to improve azimuthal resolution, utilize pulse compression technique to improve range resolution, thereby acquisition is the aperture radar more performance than really.Polarization SAR belongs to the category of SAR, and than traditional SAR, it utilizes multichannel transmitting-receiving electromagnetic wave technology, through the decipher to its different passages, can obtain target more comprehensively is familiar with.Understanding and decipher to the polarization SAR image belong to the Flame Image Process category, also relate to numerous subjects such as signal Processing, pattern-recognition and machine learning.The polarization SAR image classification is just receiving concern more and more widely as one of key link of polarization SAR Flame Image Process in national defence and civil area.Existing polarization SAR image classification method roughly can be divided into based on the method for scattering properties with based on the statistical property method, like the sorting technique of decomposing based on freeman with based on the sorting technique of scattering entropy, scattering angle etc.

Wherein, based on the polarization SAR image classification method of statistical property, be that the pixel that has certain aspect similar features in the polarization SAR image is divided into one type, there has been the sorting algorithm of a lot of maturations to be used in the polarization SAR image classification.As a new branch of science, main research work at present concentrates on the classifier design of extraction and suitable polarization SAR image classification of scattering signatures.The method for distilling of different polarization scattering signatures obtains different character, and different character is suitable for the sorter of different performance again.Existing feature extracting method mainly contains freeman and decomposes and the cloud decomposition method; Typical sorter has Wishart sorter and H/alpha sorter etc.Wherein, utilize the freeman decomposition method to obtain characteristic the polarization SAR image is carried out the branch time-like, exist the deficiency in the cognition of characteristic distribution characteristic; When utilizing the H/alpha sorter to carry out kind judging, the deficiency that exists classification judgement boundary needs artificially to confirm, these deficiencies have limited its widespread use on the polarization SAR image classification.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, proposed a kind of polarization SAR image classification method,, avoid classification judgement boundary to need the artificial problem of confirming with the distribution character of clear and definite characteristic based on the eigenvalue distribution characteristic.

Realize that the object of the invention technical scheme is through researching and analysing the distribution character of eigenwert, in conjunction with Bayes the polarization SAR image is classified, its step comprises as follows:

(1) all pixels of treating classification polarization SAR image carry out characteristic value decomposition;

(2) respectively each pixel is decomposed the eigenvalue that obtains ₁, λ ₂, λ ₃Sort, make λ ₁>=λ ₂>=λ ₃

(3) from the polarization SAR image of waiting to classify, distribute manual construction even matter regional ensemble: the C={C of k class in the polarization SAR image that waits to classify according to actual atural object ₁..., C _k, and extract the regional eigenwert of the even matter of selected k class;

(4) adopt the EM algorithm respectively to the regional eigenvalue of the even matter of k class ₁, λ ₂, λ ₃The gauss hybrid models parameter estimate, obtain the probability density function of each eigenwert in every type of even matter zone:

Wherein λ is that Mean Parameters, σ are that variance parameter, m are mixing number, the μ of Gaussian function for certain eigenwert in certain type of even matter zone, hybrid weight, the μ that ω is Gaussian function _iBe the average of i Gaussian function, σ _iBe the variance of i Gaussian function, i=1,2 ..., k;

The probability density function of each eigenwert in the every type of even matter zone that (5) utilizes step 4 to try to achieve, calculate the eigenwert joint probability density distribution function in every type of even matter zone:

$p\left(x\right|x\∈C_i)=\underset{j=1}{\overset{3}{\mathrm{\Π}}}p\left({\mathrm{\λ}}_j\right|\mathrm{\μ},\mathrm{\σ}),$

Wherein, x is even matter zone C _iIn pixel, λ _jBe j the eigenwert of pixel x, i=1,2 ..., k, j=1,2,3;

(6) according to the MAP principle, according to expression formula: x ∈ C _iSt.p (x|C _i)=maxp (x|x ∈ C _k), i=1,2 ..., k treats whole pixels in selected even matter zone in the classification polarization SAR image and classifies, and the output category result.

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

1) the present invention is owing to carried out sufficient research to the eigenvalue distribution characteristic in different even matter zone in the polarization SAR image; Express the othernesses of eigenvalue distribution characteristic between the different even matter zones with different probability density functions, the otherness between the clear and definite all kinds of even matter provincial characteristics value distribution characters;

2) the existing relatively polarization SAR image classification technology of the present invention is started with from the statistics angle, realizes classification with bayes classification method, avoids classification judgement boundary to need the artificial problem of confirming.

The simulation experiment result shows that the polarization SAR image classification method based on the eigenvalue distribution statistical property that the present invention proposes can be effectively applied to the polarization SAR image classification, and further is applied to Radar Targets'Detection and Target Recognition.

Description of drawings

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

Fig. 2 is the RGB composite diagram of the polarization SAR image of emulation employing of the present invention;

Fig. 3 is the even matter area schematic that emulation of the present invention is extracted in the polarization SAR image;

Fig. 4 is the eigenwert actual distribution and and the estimation distribution schematic diagram in the even matter of emulation of the present invention zone.

Embodiment

With reference to Fig. 1, practical implementation of the present invention is following:

All pixels that step 1. is treated classification polarization SAR image carry out characteristic value decomposition.

The polarization SAR image comprises abundant amplitude information and phase information, and the information of its each pixel is represented that by Hermitian matrix T matrix size is 3 * 3.Because therefore the information that on behalf of matrix, eigenwert can comprise select the characteristic of eigenwert as the polarization SAR image for use, utilizes the eigs function of MATLAB that the matrix T of each pixel is carried out feature decomposition, the decomposition expression formula is following:

$\left[T\right]=\left[U_3\right]\left[\begin{array}{ccc}{\mathrm{\λ}}_1& 0& 0\\ 0& {\mathrm{\λ}}_2& 0\\ 0& 0& {\mathrm{\λ}}_3\end{array}\right]{\left[U_3\right]}^{*};$

Wherein, U ₃Be the proper vector of picture element matrix T characteristic value decomposition, λ ₁, λ ₂, λ ₃Carry out the eigenwert that varies in size of characteristic value decomposition gained for picture element matrix T.

Step 2. is decomposed the eigenvalue that varies in size that obtains to each pixel respectively ₁, λ ₂, λ ₃Principle according to from big to small sorts, and makes λ ₁>=λ ₂>=λ ₃

Step 3. is chosen even matter zone, the training dataset and the test data set in the even matter of structure zone.

3a) from the polarization SAR image of waiting to classify, distribute the manual construction even matter regional ensemble of k class in the polarization SAR image of waiting to classify according to actual atural object:

C={C ₁..., C _k, k=2,3 ..., n, wherein: n waits to classify the pixel number of polarization SAR image;

3b) extract in the even matter zone of selected k class pixel as the data set of algorithm, wherein 10%～50% pixel is as the training dataset in every type of even matter zone in every type of even matter zone, and all the other are as test data set that should even matter zone.

Step 4. adopts the EM algorithm to find the solution the probability density function of each eigenwert of training dataset.

4a) with the eigenvalue of all kinds of even matter regional training data centralization pixels ₁, λ ₂, λ ₃Respectively as the input data of EM algorithm;

The mixing number m value of 4b) setting Gaussian function in the gauss hybrid models that the EM algorithm found the solution is 1, hybrid weight ω is 1, EM algorithm iteration end condition be the iteration error of algorithm less than 1.0E-06, initialization average μ and variances sigma are random value;

4c) average μ and variances sigma are upgraded through the expectation maximization flow process in the EM algorithm; The iteration error of each renewal back computational algorithm also judges whether to satisfy stopping criterion for iteration; Iteration stopping when satisfying stopping criterion for iteration, output obtains the average μ and the variances sigma of input feature vector value;

4d) will mix average μ, the variances sigma substitution formula of number m, hybrid weight ω and output:

$p\left(\mathrm{\λ}\right|\mathrm{\ω},\mathrm{\μ},\mathrm{\σ})=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i\×\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_i}\exp (-\frac{{(x-{\mathrm{\μ}}_i)}^2}{2{\mathrm{\σ}}_i^2})$

Obtain the probability density function of input feature vector value.

The joint probability function distribution and expression formula of the eigenwert in the even matter of step 5. structure zone.

The probability density function of each eigenwert in every type of even matter zone of trying to achieve with step 4, calculates every type of eigenwert joint probability density distribution function that even matter is regional:

$p\left(x\right|x\∈C_i)=\underset{j=1}{\overset{3}{\mathrm{\Π}}}p\left({\mathrm{\λ}}_j\right|\mathrm{\μ},\mathrm{\σ}),$

Wherein, x is even matter zone C _iIn pixel, λ _jBe j the eigenwert of pixel x, i=1,2 ..., k, j=1,2,3.

The pixel that step 6. pair test data is concentrated is classified the classification accuracy of calculating pixel point.

4a) according to the MAP principle, according to expression formula: x ∈ C _iSt.p (x|C _i)=maxp (x|x ∈ C _k), i=1,2 ..., k classifies to the pixel that test data is concentrated, and the output category result;

4b) test data is concentrated classification and the step 4a of pixel) in the classification of output compare, classification is identical then for classification is correct, otherwise is classification error; Calculate classification accuracy and output according to classification correct number and classification error number.

Effect of the present invention can be verified through following emulation experiment.

1. experiment condition

1.1) the polarization SAR image chosen is as shown in Figure 2 in the experiment; This figure is the RGB composograph of the L-band polarization SAR in certain zone, farmland, Dutch Flevoland area, wherein R representative | HH-VV|, and the G representative | HV|+|VH|; The B representative | HH+VV|, the picture size size is 1024 * 750.The increased surface covering in zone is farmland and waters in this polarization SAR image.

1.2) five types of even matter zones choosing are as shown in Figure 3 in the experiment, wherein Fig. 3 a, Fig. 3 b, Fig. 3 c, Fig. 3 d and Fig. 3 e are five types of different types of crops zones.

2. experiment content

2.1) each eigenwert actual distribution in above-mentioned even matter zone is experimentized with estimating to distribute, its result is as shown in Figure 4, and wherein: Fig. 4 a, Fig. 4 b and Fig. 4 c respectively are eigenvalue ₁, λ ₂, λ ₃Actual distribution with estimate distribution schematic diagram.Solid line among Fig. 4 is the actual distribution curve of eigenwert, and dotted line is the estimation distribution curve of eigenwert.Can find out eigenvalue from Fig. 4 a, Fig. 4 b and Fig. 4 c ₁, λ ₂, λ ₃Deviation between actual distribution curve and the estimation distribution curve is very little, shows that this regional eigenvalue distribution characteristic is well expressed through gauss hybrid models.

The gauss hybrid models Mean Parameters that above-mentioned five types of each eigenwerts of zone are estimated to obtain is shown in table one, and variance parameter is shown in table two:

Table one Mean Parameters

Table two variance parameter

Can find out that from table one and table two average of the gauss hybrid models parameter that the eigenwert in different even matter zone is obeyed is different with variance, show the otherness of eigenvalue distribution between the different even matter zone.

2.2) emulation classification accuracy result is shown in table three:

The classification accuracy of table three emulation experiment

Can find out that from table three the present invention is more than 90% for the classification accuracy of polarization SAR image basically, shows that the present invention can effectively solve polarization SAR image classification problem.

Claims (3)

1. based on the polarization SAR image classification method of eigenvalue distribution characteristic, may further comprise the steps:

(1) all pixels of treating classification polarization SAR image carry out characteristic value decomposition;

(2) respectively each pixel is decomposed the eigenvalue that varies in size that obtains ₁, λ ₂, λ ₃Order according to from big to small sorts, and makes λ ₁>=λ ₂>=λ ₃

(3) from the polarization SAR image of waiting to classify, distribute manual construction even matter regional ensemble: the C={C of k class in the polarization SAR image that waits to classify according to actual atural object ₁..., C _k, and extract the regional eigenwert of the even matter of selected k class;

(4) adopt the EM algorithm respectively the k class to be spared the eigenvalue of each pixel in the matter zone ₁, λ ₂, λ ₃The gauss hybrid models parameter estimate, obtain the probability density function of each eigenwert in every type of even matter zone:

Wherein λ is that Mean Parameters, σ are that variance parameter, m are mixing number, the μ of Gaussian function for certain eigenwert in certain type of even matter zone, hybrid weight, the μ that ω is Gaussian function _iBe the average of i Gaussian function, σ _iBe the variance of i Gaussian function, i=1,2 ..., m;

The probability density function of each eigenwert in the every type of even matter zone that (5) utilizes step (4) to try to achieve, calculate the eigenwert joint probability density distribution function in every type of even matter zone:

$p\left(x\right|x\∈C_i)=\underset{j=1}{\overset{3}{\mathrm{\Π}}}p\left({\mathrm{\λ}}_j\right|\mathrm{\μ},\mathrm{\σ}),$

Wherein, x is even matter zone C _iIn pixel, λ _jBe j the eigenwert of pixel x, i=1,2 ..., k, j=1,2,3;

(6) according to the MAP principle, according to expression formula: x ∈ C _n, st.p (x|C _n)=max p (x|x ∈ C _i), i=1,2 ..., k, n=1,2 ..., k treats whole pixels in selected even matter zone in the classification polarization SAR image and classifies, and the output category result.

2. polarization SAR image classification method according to claim 1, wherein described all pixels of treating classification polarization SAR image of step (1) carry out characteristic value decomposition, carry out according to following steps:

2a) will wait to classify the Hermitian matrix of polarization SAR image slices vegetarian refreshments as the input data;

2b) utilize the eigs function in the MATLAB software that the Hermitian matrix of each pixel of input is carried out characteristic value decomposition, obtain eigenwert.

3. polarization SAR image classification method according to claim 1, wherein the described employing of step (4) EM algorithm is spared the eigenvalue of each pixel in the matter zone respectively to the k class ₁, λ ₂, λ ₃The gauss hybrid models parameter estimate, carry out according to following steps:

3a) with the eigenvalue of each pixel in all kinds of even matter zone ₁, λ ₂, λ ₃Respectively as the input data of EM algorithm;

The mixing number m value of 3b) setting Gaussian function in the gauss hybrid models that the EM algorithm found the solution is 1, hybrid weight ω is 1, EM algorithm iteration end condition be the iteration error of algorithm less than 1.0E-06, initialization average μ and variances sigma are random value;

3c) average μ and variances sigma are upgraded through the expectation maximization flow process in the EM algorithm; The iteration error of each renewal back computational algorithm also judges whether to satisfy stopping criterion for iteration; Iteration stopping when satisfying stopping criterion for iteration, output obtains the average μ and the variances sigma of input feature vector value;

3d) will mix average μ, the variances sigma substitution formula of number m, hybrid weight ω and output:

$p\left(\mathrm{\λ}\right|\mathrm{\ω},\mathrm{\μ},\mathrm{\σ})=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i\×\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_i}\exp (-\frac{{(x-{\mathrm{\μ}}_i)}^2}{2{{\mathrm{\σ}}_i}^2})$

Obtain the probability density function of input feature vector value.

Images