CN104463213A - Polarization SAR image classification method based on wavelet kernel sparsity LSSVM - Google Patents

Abstract

The invention discloses a polarization SAR image classification method based on a wavelet kernel sparsity LSSVM. The method mainly solves the problems that when polarization SAR images are classified in the prior art, consumed time is long and the classification precision is low. The method comprises the realization steps that 1), coherence matrixes of the polarization SAR images are input and filtered; 2), the filtered coherence matrixes are drawn to form a feature matrix; 3), 5% of feature vectors are selected from the feature matrix to form training samples; 4), Morlet wavelet kernel functions are selected to map the training samples to a high-dimensional space; 5), the training samples in the high-dimensional space and class labels corresponding to the training samples are utilized for training to form a wavelet kernel sparsity classifier; 6), the classifier is utilized for classifying the polarization SAR images. The method has the advantages that calculation complexity is low, the fitting effect for the optimal classification functions is good, the classification speed is high, the classification precision is high, and the method can be used for recognizing, tracking and locating targets.

Description

Based on the Classification of Polarimetric SAR Image method of the sparse LSSVM of Wavelet Kernel

Technical field

The invention belongs to technical field of image processing, further relate to a kind of polarimetric synthetic aperture radar SAR image sorting technique based on Wavelet Kernel Sparse least squares support vector machine LSSVM in Image Classfication Technology field, can be used for carrying out terrain classification to Polarimetric SAR Image and marking, realize target recognition and tracking is located.

Background technology

Classification of Polarimetric SAR Image can think the process of classifying to the pixel in image, judges the classification belonging to this pixel, complete the classification to image according to the attributive character of pixel each in image.Because Statistical Learning Theory is a kind of machine Learning Theory specialized under Small Sample Size, the support vector machine of the Corpus--based Method theories of learning is a kind of effective supervised classifiers, has been widely used in the fields such as target identification and Iamge Segmentation.

The paper " application of LSSVM algorithm in polarization SAR impact classification " of Meng Yun shwoot table (" geospatial information ", number: disclose the method for a kind of LSSVM to Classification of Polarimetric SAR Image 1672-4623 (2012) 03-0043-03) by article.The implementation procedure of the method is, first goal decomposition is carried out to polarization SAR image, extract the vector set of 5 parameter compositions as feature, secondly feature vector set is carried out linear normalization process, finally traditional SVM classifier and LSSVM sorter are carried out performance comparison, and adopt LSSVM to obtain classification results.The weak point of the method is, the method have finally chosen LSSVM sorter, and this disaggregated model cannot ensure that the solution obtained is globally optimal solution, and it is openness to separate shortage, easily cause over-fitting, the impact of isolated point and noise cannot be overcome, cause nicety of grading low.

Patent " a kind of Classification of Polarimetric SAR Image method based on semi-supervised SVM and MeanShift " (number of patent application: 201410076676.9, the publication No.: CN 103914704A) of Xian Electronics Science and Technology University's application.The implementation procedure of the method is, first Classification of Polarimetric SAR Image training set and class test collection is set up respectively, secondly the classification results of Polarimetric SAR Image is obtained with SVM algorithm, choose the sample set that degree of confidence is high, svm classifier result is revised with MeanShift, upgrade sample set, test set and disaggregated model, finally with this disaggregated model, Polarimetric SAR Image is classified.The weak point of the method is, the method have employed traditional SVM classifier, need to ask disaggregated model by quadratic programming, so when classifying to Polarimetric SAR Image, because data volume is comparatively large, cause computation complexity high, training time is long, inefficiency, and poor to the fitting effect of complicated function, cause nicety of grading lower.

Summary of the invention

The object of the invention is to propose a kind of Classification of Polarimetric SAR Image method based on Wavelet Kernel Sparse least squares support vector machine LSSVM, to shorten the classification time, improve nicety of grading, realize the quick and precisely classification to polarimetric synthetic aperture radar SAR image.

For achieving the above object, technical scheme of the present invention comprises the steps:

(1) coherence matrix that a width size is the Polarimetric SAR Image to be sorted of 3*3*N is inputted, and adopt the Lee wave filter filtering coherent noise of 7*7 window size, obtain filtered coherence matrix T, wherein, in filtered coherence matrix T, each element is a 3*3 matrix, and N represents the sum of polarimetric synthetic aperture radar SAR image pixel to be sorted;

(2) by 3*3 matrix corresponding for element each in coherence matrix T after filtering, pull into the proper vector of a 1*9, form an eigenmatrix T ¹;

(3) from eigenmatrix T ¹the proper vector of middle Stochastic choice 5% is as training sample set X, and remaining 95% as test sample book collection Xt;

(4) choose Morlet Wavelet Kernel Function, and with this Wavelet Kernel Function, training sample set X is mapped to higher dimensional space from luv space, obtain the training sample set X in higher dimensional space ¹.

(5) by the training sample set X in higher dimensional space ¹be input in Sparse least squares support vector machine LSSVM sorter with the class label collection of its correspondence, train Wavelet Kernel Sparse least squares support vector machine LSSVM sorter;

(6) with Wavelet Kernel Sparse least squares support vector machine LSSVM sorter, each pixel in polarimetric synthetic aperture radar SAR image to be sorted is marked, obtains classification results, complete classification.

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

The first, computation complexity is low, and classification speed is fast.

Support vector machine LSSVM sorter is supported owing to present invention employs sparse least squares, solve least square method supporting vector machine LSSVM sorter in prior art and need to solve the complex model of multiple linear equation, cause computation complexity high, the solution obtained is not sparse, cause the problem of classification duration, achieve the Fast Classification to Polarimetric SAR Image.

The second, nicety of grading is high.

The present invention is owing to being improved to Morlet Wavelet Kernel Function by the inner core function of sparse LSSVM sorter, solving SVM classifier in prior art adopts the fitting effect of Radial basis kernel function to complicated function poor, cause the problem that nicety of grading is lower, make to adopt technology in the present invention when to Classification of Polarimetric SAR Image, under the prerequisite of not sacrificing classification effectiveness, improve nicety of grading.

Accompanying drawing explanation

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

Fig. 2 is the Flevoland of AIRSAR platform acquisition in 1989, the pcolor looking polarization SAR Data Synthesis figure of the L-band in Netherlands area more;

Fig. 3 is the Flevoland of AIRSAR platform acquisition in 1989, the figure of substance markers practically looking polarization SAR Data Synthesis figure of the L-band in Netherlands area more;

Fig. 4 is the result schematic diagram that the present invention classifies to Fig. 2.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further detail.

With reference to Fig. 1, as follows to implementation step of the present invention:

Step 1, input data.

In embodiments of the present invention, the coherence matrix that a width size is the Polarimetric SAR Image to be sorted of 3*3*N is inputted by WINDOWS XP system, and adopt the Lee wave filter filtering coherent noise of 7*7 window size, obtain filtered coherence matrix T, wherein, in filtered coherence matrix T, each element is a 3*3 matrix, and N represents the sum of polarimetric synthetic aperture radar SAR image pixel to be sorted.

Step 2, feature extraction.

By 3*3 matrix corresponding for element each in coherence matrix T after filtering, pull into the proper vector of a 1*9, form an eigenmatrix T ¹.

Step 3, chooses training sample.

From eigenmatrix T ¹the proper vector of middle Stochastic choice 5% is as training sample set X, and remaining 95% as test sample book collection Xt, and wherein training sample concentrates each training sample to be the proper vector of a 1*9.

Step 4, chooses kernel function.

4.1) condition choosing Wavelet Kernel Function is set: (a) meets the structural environment of support vector machine kernel function; B () computation complexity can not be too high, too much parameter can strengthen the training time;

4.2) as follows according to choosing the Morlet Wavelet Kernel Function that condition chooses:

$K(x_i,x_j)=\underset{k=1}{\overset{l}{\mathrm{\Π}}}\left(\cos (1.75\×\frac{x_i^k-x_j^k}{a})\exp (-\frac{{\left|\right|x_i^k-x_j^k\left|\right|}_2^2}{{2a}^2})\right)$

Wherein, K (x _i, x _j) represent training sample x in training sample set X _iand x _jcore, i, j ∈ [1, n ₀], n ₀represent the number of training sample in training sample set X, l represents the dimension of training sample set X, and a represents Morlet Wavelet Kernel parameter, represent training sample x respectively _i, x _jkth dimension value, represent training sample x _iand x _jkth dimension between Euclidean distance square;

4.3) with the Morlet Wavelet Kernel Function K (x chosen _i, x _j) training sample set X is mapped to higher dimensional space from luv space, obtain the training sample set X in higher dimensional space ¹.

Step 5, training classifier.

By the training sample set X in higher dimensional space ¹be input in Sparse least squares support vector machine LSSVM sorter with the class label collection of its correspondence, train Wavelet Kernel Sparse least squares support vector machine LSSVM sorter.

What described Sparse least squares support vector machine LSSVM sorter adopted is people such as the Bao Liefeng Sparse least squares support vector machine LSSVM tool box in the paper " Fast Sparse Approximation for Least Squares Support VectorMachine " to deliver for 2009, the expression of the Wavelet Kernel Sparse least squares support vector machine LSSVM sorter trained is as follows:

$f\left(t\right)=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i\underset{j=1}{\overset{l}{\mathrm{\Π}}}\cos \left(1.75\frac{t^j-s_i^j}{a}\right)\exp (-\frac{1}{2}{\left(\frac{t^j-s_i^j}{a}\right)}^2)+b)$

Wherein, t represents sample to be sorted, and f (t) represents the classification results of sample t to be sorted, and n expresses support for vectorial number, α _irepresent i-th support vector s _icorresponding coefficient, y _irepresent i-th support vector s _icorresponding class label, l represents the dimension of sample t to be sorted, and a represents Morlet Wavelet Kernel letter parameter, t ^jrepresent the jth dimension value of sample t to be sorted, represent i-th support vector s _ijth dimension value, b represents offset parameter.

Step 6, with Wavelet Kernel Sparse least squares support vector machine LSSVM sorter, marks each pixel in polarimetric synthetic aperture radar SAR image to be sorted, obtains classification results, complete classification.

The present invention can be verified by follow-up emulation experiment.

1, emulation experiment condition:

This emulation experiment needs to choose a width and has the Polarimetric SAR Image of substance markers figure practically as experimental image, the experimental image chosen is the Flevoland that AIRSAR platforms in 1989 as shown in Figure 2 obtain, the pcolor looking polarization SAR Data Synthesis figure of the L-band in Netherlands area more, picture size size is 750 pixel × 1024 pixels, have 11 atural object classifications, substance markers figure is Fig. 3 practically; This emulation experiment hardware platform is: Intel Core2Duo CPU i3@3.2GHZ, 3GB RAM, software platform: MATLAB R2010a.

2, emulation experiment content and interpretation of result:

Emulation experiment 1, adopt the present invention to classify to Fig. 2, result as shown in Figure 4.

As can be seen from Figure 4, except Wheat atural object and object area wrong branch in Potatoes ground is more, result fuzzyyer except, remaining region all obtains more accurate classification results, and the edge of view picture result figure is smoother, clear and legible.As can be seen here, technical method of the present invention is applicable to carry out terrain classification to Polarimetric SAR Image, and can obtain good classifying quality.

Emulation experiment 2, SVM algorithm in the present invention and prior art and sparse LSSVM algorithm are carried out the contrast of classification time and nicety of grading, comparing result is as shown in table 1, the classification time in table 1 is sorter from beginning label to completing the classification time used, the ratio of nicety of grading shared by pixel number identical with label in the mark result shown in Fig. 3 in the classification results that obtains.

Table 1 three kinds of classification Comparative result tables

As can be seen from Table 1, adopt the present invention, when classifying to Polarimetric SAR Image, from nicety of grading, total nicety of grading of the present invention apparently higher than employing SVM algorithm and sparse LSSVM algorithm, and all obtains good classifying quality to all atural object classifications; From the classification time, the present invention's shortest time used, and compare SVM algorithm, effect is very obvious.

To sum up, when adopting the present invention to classify to Polarimetric SAR Image, classification speed and nicety of grading all increase, and demonstrate validity of the present invention further.

Claims (3)

1., based on a Classification of Polarimetric SAR Image method of the sparse LSSVM of Wavelet Kernel, concrete steps are as follows:

(1) coherence matrix that a width size is the Polarimetric SAR Image to be sorted of 3*3*N is inputted, and adopt the Lee wave filter filtering coherent noise of 7*7 window size, obtain filtered coherence matrix T, wherein, in filtered coherence matrix T, each element is a 3*3 matrix, and N represents the sum of polarimetric synthetic aperture radar SAR image pixel to be sorted;

(2) by 3*3 matrix corresponding for element each in coherence matrix T after filtering, pull into the proper vector of a 1*9, form an eigenmatrix T ¹;

(3) from eigenmatrix T ¹the proper vector of middle Stochastic choice 5% is as training sample set X, and remaining 95% as test sample book collection Xt;

(4) choose Morlet Wavelet Kernel Function, and with this Wavelet Kernel Function, training sample set X is mapped to higher dimensional space from luv space, obtain the training sample set X in higher dimensional space ¹;

(5) by the training sample set X in higher dimensional space ¹be input in Sparse least squares support vector machine LSSVM sorter with the class label collection of its correspondence, train Wavelet Kernel Sparse least squares support vector machine LSSVM sorter;

(6) with Wavelet Kernel Sparse least squares support vector machine LSSVM sorter, each pixel in polarimetric synthetic aperture radar SAR image to be sorted is marked, obtains classification results, complete classification.

2. the Classification of Polarimetric SAR Image method based on the sparse LSSVM of Wavelet Kernel according to claim 1, is characterized in that, the Morlet Wavelet Kernel Function chosen in described step (4), and computing formula is as follows:

$K(x_i,x_j)=\underset{k=1}{\overset{l}{\mathrm{\Π}}}\left(\cos (1.75\×\frac{x_i^k-x_j^k}{a})\exp (-\frac{{\left|\right|x_i^k-x_j^k\left|\right|}_2^2}{{2a}^2})\right)$

Wherein, K (x _i, x _j) represent training sample x in training sample set X _iand x _jcore, i, j ∈ [1, n ₀], n ₀represent the number of training in training sample set X, l represents the dimension of training sample, and a represents Morlet Wavelet Kernel parameter, represent training sample x respectively _i, x _jkth dimension value, represent training sample x _iand x _jkth dimension between Euclidean distance square.

3. the Classification of Polarimetric SAR Image method based on the sparse LSSVM of Wavelet Kernel according to claim 1, is characterized in that, the Wavelet Kernel Sparse least squares support vector machine LSSVM sorter in described step (5), and it is expressed as follows:

$f\left(t\right)=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i\underset{j=1}{\overset{l}{\mathrm{\Π}}}\cos \left(1.75\frac{t^j-s_i^j}{a}\right)\exp (-\frac{1}{2}{\left(\frac{t^j-s_i^j}{a}\right)}^2)+b)$

Wherein, t represents sample to be sorted, and f (t) represents the classification results of sample t to be sorted, and n expresses support for vectorial number, α _irepresent i-th support vector s _icorresponding coefficient, y _irepresent i-th support vector s _icorresponding class label, l represents the dimension of sample t to be sorted, and a represents Morlet Wavelet Kernel parameter, t ^jrepresent the jth dimension value of sample t to be sorted, represent i-th support vector s _ijth dimension value, b represents offset parameter.