Texture Feature Extraction Method Combining ... - Semantic Scholar
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
675
Texture Feature Extraction Method Combining Nonsubsampled Contour Transformation with Gray Level Co-occurrence Matrix He Xiaolan and Wu Yili Business School, Zhejiang University City College, Hangzhou 310015, China
Wu Yiwei Valparaiso University, IN 46383, USA
Abstract—Gray level co-occurrence matrix (GLCM) is an important method to extract the image texture features of synthetic aperture radar (SAR). However, GLCM can only extract the textures under single scale and single direction. A kind of texture feature extraction method combining nonsubsampled contour transformation (NSCT) and GLCM is proposed, so as to achieve the extraction of texture features under multi-scale and multi-direction. We firstly conducted multi-scale and multi-direction decomposition on the SAR images with NSCT, secondly extracted the symbiosis amount with GLCM from the obtained sub-band images, then conducted the correlation analysis for the extracted symbiosis amount to remove the redundant characteristic quantity; and combined it with the gray features to constitute the multi-feature vector. Finally, we made full use of the advantages of the support vector machine in the aspects of small sample database and generalization ability, and completed the division of multifeature vector space by SVM so as to achieve the SAR image segmentation. The results of the experiment showed that the segmentation accuracy rate could be improved and good edge retention effect could be obtained through using the GLCM texture extraction method based on NSCT domain and multi-feature fusion in the SAR image segmentation. Index Terms—Synthetic Aperture Radar; Image Segmentation; Nonsubsampled Contour Transformation; Gray Level Co-Occurrence Matrix; Support Vector Machine; Feature Selection
I.
INTRODUCTION
Synthetic aperture radar (SAR) can form images at all time and all weather, and it plays an increasingly significant role in the military and civilian aspects. Therefore, each country pays great attention to the research and application of the relevant technology of SAR. SAR image interpretation is a very important research field of SAR, while SAR image segmentation is the key pretreatment process in the SAR image interpretation. Because the segmentation quality of the images directly influence the late target recognition of SAR images (Such as airports, ports, roads and bridges), the image segmentation is viewed as the bottleneck in the SAR image interpretation [1]. The general image
© 2013 ACADEMY PUBLISHER doi:10.4304/jmm.8.6.675-684
segmentation methods often fail due to the imaging particularity of SAR images. Therefore, it is very necessary to research effective SAR image segmentation algorithm. At present SAR image segmentation is roughly divided into thresholding method, edge segmentation method, statistical segmentation method and the segmentation method based on pixel features [2]. The thresholding method, edge segmentation method, and statistical segmentation method all use the gray information to achieve the segmentation. However, it is difficult to achieve good SAR image segmentation just by using gray features. The reason is because SAR images reflect the backscattering characteristics of ground objects on the radar wave. If different ground objects have same or similar backscattering coefficients, they will exhibit same or similar gray values in the SAR images, and accordingly confusion will occur. The texture information of SAR images is very rich. The different shapes and different physical features of the imaging targets show different textures, and these textures can become the important basis for the radar to distinguish different objects. The segmentation method based on pixel features can make full use of the intrinsic properties of the images (Such as texture features), and conduct multi-feature fusion through extracting various different features, so as to achieve high segmentation precision. Gray level co-occurrence matrix (GLCM) is the common method to describe the textures. However, its descriptions of the texture features are not so detailed that it is not good enough to extract the features by this method. In order to make full use of the advantages of GLCM and multi-resolution transformation, many scholars [4] have researched the texture feature extraction method of multi-scale GLCM. Su Hui [5] proposed a method to conduct the symbiosis feature extraction on the approximation sub-band on each scale of wavelet transform. This method can describe the texture structure effectively under different resolutions and it is successfully used for the fabric texture recognition, but it ignores to use the detail sub-band of wavelet transform. Han Yanfang [6] proposed to use GLCM in the wavelet
676
detail sub-band to extract the symbiosis features of the texture, and used it for the defect detection of texture surface. But it is insufficient to describe different texture details just with wavelet detail sub-band. Shang Zhaowei [7] et al. extracted 7 symbiosis features of coefficient sub-band in the complex wavelet transform domain and used them for texture retrieval, but no consideration was given to the problems of parameter optimization and feature selection when extracting the symbiosis features. In view of the above problems, Zhong Hua [8] et al. proposed the multi-resolution co-occurrence matrix feature extraction method through extracting the cooccurrence matrix on the approximation sub-band and detail sub-band of the nonsubsampled wavelet transform simultaneously, so as to obtain good classification results in the classification of texture images. Clausi [9] et al. comprehensively used the airspace gray level cooccurrence matrix and Gabor wavelet to extract the texture features of SAR images, and thus conducted SAR sea ice image segmentation. The experiments showed that the classification accuracy of using different ways to extract the texture feature for fusion was higher than that of the classification only using a kind of texture. Wang Haifeng [10] et al. proposed a kind of texture image segmentation algorithm combining nonsubsampled contour transformation (NSCT) and support vector machine (SVM), used NSCT to decompose the images, then extracted the texture features, and finally completed the segmentation of the entire feature image by SVM. However, only the multi-scale feature of the image is extracted through this method, while the spatial structure information of the texture is ignored. In view of the features such as large gray change, complex texture and fuzzy boundaries, Wang Qingxiang [11] et al. proposed a kind of unsupervised segmentation method of SAR images based on multi-features, and obtained good segmentation effect. The several feature extraction methods mentioned above are all conducted in the airspace or wavelet domain. However, the wavelet transform has limited isotropy and directionality (Three), there is no translation invariant, and the singularity on the line is not the optimal basis, so the texture information is not fully shown. As a kind of multi-scale geometric analysis tool, NSCT has the optimal representation form on the curve singularity function, it can overcome the above-mentioned shortcomings of the wavelet transform, and it has obtained good effect in the field of the image analysis (such as noise reduction, feature extraction and retrieval). GLCM is a kind of texture analysis method which is used most widely and it is commonly used to extract the texture features of the images. The multi-resolution and multi-direction characteristics of NSCT can be combined with the texture structure information of GLCM to extract the texture features of SAR images with GLCM in the NSCT domain, so that the gray symbiosis features can provide the dynamic information of texture space scale change and have better descriptive ability for the texture. At the same time, we can give full play to the characteristics of multi-direction of NSCT by extracting
© 2013 ACADEMY PUBLISHER
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
the symbiosis amount of different sub-band in the NSCT domain (Each sub-band of NSCT represents different direction information.). In this paper, we have researched the GLCM texture feature extraction method based on NSCT domain. The major innovation points are as follows: 1. The anisotropy, multi-resolution and multidirection characteristics of NSCT are used, multiscale and multi-direction decompositions are conducted for the SAR images by using NSCT, the sub-band images of different scales and different directions are obtained, the high-dimension singularity information and rich texture information in the images are captured, and the limitations of the wavelet transform are overcome. 2. The spatial distribution information and structure information property of the images can be described using GLCM. NSCT is combined with GLCM, and GLCM is used for the sub-band images of different scales and different directions after passing NSCT to extract the gray symbiosis feature quantity, so as to overcome the limitations that wavelet domain GLCM can only extract the gray symbiosis amount in the limited directions and it cannot describe the line singularity effectively, and thus achieve the extraction of the texture features of multi-resolution and multi-direction. 3. In this method, the extracted gray values and gray variances are taken as the gray features, the multifeature fusion strategy (namely combining texture feature and gray features) is used, so as to constitute the multi-feature values and use them for the SAR image segmentation. 4. Since the support vector machine (SVM) has higher classification accuracy than other classifiers in the aspects such as high-dimension feature vector and small sample data, in this method SVM is used to complete the division of high-dimension feature space so as to achieve the purpose of segmenting SAR images. The experimental results show that the GLCM feature extraction method based on NSCT domain is an effective method to extract the features of SAR images and its segmentation accuracy rate exceeds that of GLCM feature extraction method of wavelet domain. This method can improve the accuracy of the segmentation image edge location and reduce the misclassification of the pixels within the region, so as to obtain good segmentation results. II.
GLCM FEATURE EXTRACTION METHOD BASED ON NSCT DOMAIN
A. Nonsubsampled Contour Transformation Nonsubsampled Contour Transformation [12] (NSCT) is a kind of algorithm which uses the iterative nonsubsampled filter bank to achieve a series of multiresolution, multi-direction and translation invariant frequency domain sub-images. In this algorithm, multiscale analysis is conducted firstly on the images, then multi-direction analysis is conducted, and the layer by
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
layer iteration is conducted separately for these two steps. The structure of NSCT is divided into nonsubsampled pyramid filter bank (NSPFB) and nonsubsampled directional filter bank (NSDFB), and its realization structural principle is as shown in Fig. 1. Its decomposition process can be expressed as: Firstly the nonsubsampled pyramid filter bank (NSPFB) is used to conduct multi-scale decomposition on the images, so as to obtain various high-frequency sub-band images with different frequencies and a low-frequency sub-band image, and then the nonsubsampled directional filter bank is used to conduct multi-direction decomposition for the obtained various high-frequency sub-band images, so as to obtain the sub-band images (coefficients) of different scales and different directions. Low pass subband Second stage low pass filter
High-frequency directional sub-band
Image
677
pixels in y direction is Ny, and the highest gray level of the images is Ng level.
Figure 2. Schematic diagram of gray level co-occurrence matrix structure
Starting from the pixel with a gray value of i , the gray value occurring simultaneously in another pixel with a straight-line distance of d and an angle of is j ; based on this, the probability of these two gray values occurring simultaneously in the entire images can be defined as: P(i, j, ) {( x, y) | f ( x, y) i, f ( x d cos , y d sin ) } (1)
High-frequency First stage low pass filter directional sub-band
Nonsubsampled pyramid Nonsubsampled directional filter bank filter bank
Figure 1. NSCT realization structural principle diagram
When the decomposition level of NSPFB is L, 1 low pass sub-band image and L high pass sub-band images can be obtained through decomposition. Because there is no down-sampling, all the sub-band images have the same sizes with the original images. The realization process of NSDFB is that firstly up-sampling is conducted on the filter, and secondly the wave filtering is conducted on the decomposed sub-band images in the two channel direction of last stage, so as to achieve the more accurate directional decomposition in the frequency domain. 2t directional sub-band images which have the same size as the original images can be obtained through l stage directional decomposition of NSDFB on a scale sub-band image. NSPFB is combined with NSDFB to achieve NSCT. The band-pass sub-band images, which are obtained after decomposing the original images through NSPFB, are inputted NSDFB and the band-pass directional sub-band images can be obtained, so as to achieve the multi-scale and multi-direction decomposition of images. Gray level co-occurrence matrix (GLCM) [13] is recognized as an important texture analytical method today. Its definition is as follows: In Fig. 2, oxy is set as the coordinate plane of the image pixel, the gray coordinate is z axis, the total number of pixels in x direction is Nx, the total number of
© 2013 ACADEMY PUBLISHER
The probability of two pixel gray levels occurring simultaneously converts the spatial coordinates of (x, y) to the gray descriptions of (i, j) and forms the gray level co-occurrence matrix. In the actual texture analysis, the gray level co-occurrence matrix often is not used directly, while the second degree statistics are extracted on the basis of this matrix and used as the feature values of texture identification. A large number of experiments showed that for the SAR images the following four statistics could best express the texture features of SAR images [14][15]. Homogeneous area:
HOM p(i, j ) / [1 (i j )2 ] i
(2)
j
Angular second moment:
ASM [ p(i, j )]2
(3)
ENT p(i, j ) log( p(i, j ))
(4)
i
j
Entropy: i
j
Dissimilarity:
DIS | i j | p(i, j ) i
III.
(5)
j
PRINCIPLES OF SUPPORT VECTOR MACHINE
The principle of the support vector machine (SVM) is that we use the classification hyper plane H: w, x b 0 to correctly separate the two sample points {( x1 Rn , yi {1, 1}, i 1, 2,.., N ) in the space, H1:
w, x b 1 , H2: w, x b 1 , as shown in Fig. 3, and we obtain the minimum distances of the positive
678
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
sample and negative sample to the hyper plane, so as to make the interval between the two samples M arg in 2 / w H2 H1
H
the high-dimension feature space can be obtained as follows:
f ( x) sgn(k ( w, x) b) sgn yi ai K ( x, xi ) b iSV
Support vector
(9)
In the practical application, the commonly used kernel functions are as follows: Linear kernel function (linear):
K ( xi , x j ) xi x j
Margin=2/ w
(10)
Radial basis kernel function (RBF): Figure 3. Schematic diagram of optimal classification hyper plane
The original problem is mathematically described as a constrained nonlinear programming problem: 1 2 Minimize ( w, b) w 2
yi ( xi w b) 1 0 i 1, 2,
,N
(6)
The minimum norm W meeting the constraint is the normal vector of the optimal classification hyper plane. The decision function based on the optimal classification hyper plane is as follows:
f ( x) sgn yi ai x, xi b iSV
(7)
In the formula, SV is the support vector, and ai 0 is the unique solution of the above-mentioned secondary programming problem. As for the nonlinear case, we just need to replace the dot product in the dual problem with the convolution kernel function K ( xi , x j ) . As for the case that the sample points cannot be separated, we can introduce the slack variable i 0(i 1, , N ) and penalty factor C 0 , where, C is used for controlling the punishment degree for the misclassified samples. The soft edge classification surface is constructed so as to allow the misclassified samples to exist. The mathematical model of the corresponding optimization problem is as follows: N 1 2 Minimize ( w, b) w C i 2 i 1
yi ( xi w b) 1 i i 0, i 1, 2,
,N
(8)
From the statistical learning theory we can obtain that, the function K ( x, y) meeting the Mercer Theorem can be used as the inner product, and the nucleation function
K ( x, y ) i i ( x) i ( y) can be expanded with the i 1
coefficient i 0 . After the kernel function is introduced, when the sample is not linearly separable in the highdimension feature space, the decision function of the generalized optimal classification hyper plane based on
© 2013 ACADEMY PUBLISHER
2
K ( xi , x j ) exp{ xi x j }
(11)
Polynomial kernel function (polynomial):
K ( xi , x j ) [( xi x j ) ]d }
(12)
Two-layer neural network kernel function (Sigmoid tanh):
K ( xi , x j ) thah(u( xi x j ) c) IV.
(13)
EXPERIMENT PROCESS
A. GLCM Texture Feature Extraction and Selection Based on NSCT Domain Assuming that the input image is f ( x, y ) and the size is M*N, the decomposition process of NSCT can be expressed as: J
lj
f ( x, y ) NSCTaJ b j , k
(14)
j 1 k 1
where, aJ is the low-frequency sub-band, and b j , k is the high-frequency sub-band of scale j and direction k. A NSCT transform coefficient can be expressed with b( j, k , m, n) , where, J is the decomposition level of NSPFB, is the decomposition level of NSDFB of layer j; j represents the scale label after the decomposition of NSPFB, k represents the direction label in NSDFB, and m and n represent the spatial location information of the coefficients in the directional sub-bands. GLCM is applied to decompose NSCT so as to obtain the lowfrequency sub-band aJ and the high-frequency sub-band
b j , k , the homogeneous area (HOM), angular second moment (ASM), entropy (ENT) and dissimilarity (DIS) of the gray symbiosis amount are extracted, the extracted symbiosis amounts are taken as the texture feature vectors, and thereby the GLCM texture feature extraction based on NSCT domain is completed. The extraction steps of the GLCM texture feature quantity based on NSCT domain are as follows: 1) NSCT decomposition of images. Firstly we use NSCT to conduct two-layer decomposition for the images. The first scale direction number is 8, the second scale direction number is 4, and
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
13 sub-bands with the same size as the original images are obtained. In this way, the images are decomposed into the sub-bands with different scales and different directions, and each directional sub-band represents a direction of texture. 2) Transform coefficient quantification. NSCT low-frequency sub-band approximately obeys the uniform distribution, while the high-frequency subband shows a distribution of “High Peak, Long Tail”, as shown in Fig. 4 (The horizontal axis represents the amplitude, and the vertical axis represents the frequency.). The NSCT coefficients and wavelet coefficients have similar statistical distribution, and thus the coefficients are quantified according to the quantification strategy of the literature [8]. The uniform quantification is applied to the low-frequency sub-bands and the quantification level is 16, that is to say, the quantification is conducted according to the formula (15), where, [] represents rounding down.
G [(C Cmin ) / (Cmax Cmin )] 16
(15)
The non-uniform quantification is applied to the detail sub-band, and the coefficient is quantified to 16 levels. Each sub-band variance l of the same scale is estimated firstly, is taken as the boundary during the quantification, and then the piecewise quantification is conducted according to formula (16).
Gl
[( C Cmin )/( Cmax Cmin )]12,|C ( m , n ) 3 l [( C Cmin )/( Cmax Cmin )] 2,|C ( m, n ) 3 l
(16)
679
we applied windows of N=5×5 for sub-band images in second scale and low-frequency sub-bands. 4) Distance parameter d. Distance d reflects the frequency period of texture. The number of direction change varies in accordance with different SNCT decomposition parameter settings, and there is certain difference between different dimensions with the change of direction parameters. Therefore, different distance intervals d should be applied for different scales when calculating the gray symbiosis amount for NSCT decomposition sub-band. A small distance is applied for rough texture (high-frequency) and a large distance is applied for delicate texture (low-frequency sub-band). According to this, we applied d=1 for changing 8 subbands in the first scale and applied d=2 for changing 4 sub-bands in the second scale and low-frequency subbands. 5) Selection of direction . We drew the symbiosis amounts through GLCM formula (2) to (5) and calculated the homogeneous area, angular second moment, entropy and dissimilarity in four angles ( were respectively 0 , 45 , 90 and 135 ) of the NSCT sub-bands, so as to prepare for the following multi-feature fusion. Then we took the average of values in four directions of each symbiosis amount, as shown in formula (17), so as to improve the robustness of gray symbiosis over directions.
f ( f o f 45 f 90 f 135 ) / 4
At last, we got the gray symbiotic feature quantity of the image at point (m, n) in NSCT domain according to the parameter selection principles. We described the textural features of this point with 52-dimensional symbiosis amount:
N _ G(m, n) [ HOM i , AMSi , ENTi , DISi ]i 1,
Figure 4. NSCT coefficient gray statistical histogram
3) Window traversal method and window size are considered when selecting window. In the paper, we applied overlaid windows for window traversal images, that is, we get a small window, by regarding each pixel as the center, the characteristics calculated in the window are only the features of the center point, and the final categorical attributes are only those of the center point. The selection of window size is also critical, since a too large window will lead to large calculation amount and memory space while a too small window can’t contain complete texture information. Theoretically, large windows are applied for low-frequency rough texture; small windows are applied for high-frequency delicate texture. According to the criteria for window size selection, we applied windows of different sizes for different NSCT decomposition scales. We applied windows of N=3×3 for sub-band images in the first scale;
,13
(18)
However, there is certain correlation between the four gray symbiosis amounts as homogeneous area, angular second moment, entropy and dissimilarity. We first conducted correlation analysis over the feature quantity extracted in terms of features selection, and then we eliminated redundant features to reduce the dimensions and calculation amount. Fig. 5 shows the SAR image after noise reduction. We drew the textural features of SAR image in the same way as GLCM feature extraction based on NSCT domain, and then we divided the four textural features as homogeneous area, angular second moment, entropy and dissimilarity by SVM and got the single-feature segmentation image as shown in Fig. 6(a) to Fig. 6(d).
Figure 5. SAR image after noise reduction © 2013 ACADEMY PUBLISHER
(17)
680
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
HOM, ASM and ENT, and the correlation coefficients are all larger than 0. 6. From Table I and Fig. 6(a) to 6(b) we got that we could reduce the feature dimensions from the original 52 dimensions to 26 dimensions by just selecting two statistics amounts of HOM and ASM, SN _ G(m, n) [HOM i (m ,n ), AMSi (m ,n )],i 1, ,13 namely. (a) ASM
(b) HOM
(c) ENT
(d) DIS
Figure 6. Segmentation images based on single textural features
It can be seen from Fig. 6(a) to (d) that Fig. 6(a), Fig. 6(c) and Fig. 6(d) are very similar, i. e. , there are strong correlations between the three feature quantities (ASM, ENT and DIS). Due to certain correlations between HOM, ASM, ENT, DIS, if we conduct feature selection, we should eliminate feature quantities with strong correlation from the original feature quantities and keep those complementary ones contributive to following segmentation, so as to eliminate redundant features, reduce calculation amount and improve segmentation accuracy. Formula (19) is used to calculate the correlation between two images:
r
( A
mn
m
A)( Bmn B )
n
( ( Amn A) 2 )( ( Bmn B ) 2 ) m
n
m
(19)
n
where, A mean2( A) ; B mean2( B) ; mean() means to take the mean value of the matrix. We calculated the absolute values of correlation coefficient between any two of the above four symbiosis statistics amounts in the 13 sub-bands respectively corresponding to the same scale and the same direction. Then we took the average of the 13 sub-bands results and got the correlation coefficient table (Table I). TABLE I.
CORRELATION COEFFICIENT BETWEEN STATISTICS AMOUNTS
HOM ASM ENT DIS
HOM 1 0. 4796 0. 6282 0. 9019
ASM 0. 4796 1 0. 8649 0. 6184
ENT 0. 6282 0. 8649 1 0. 7713
DIS 0. 9019 0. 6184 0. 7713 1
It can be seen from Table I that the correlation coefficient between statistics amounts HOM and ASM is smaller than 0. 5; the correlation coefficient between HOM and ENT and the correlation coefficient between HOM and DIS are both larger than 0. 6, while the correlation coefficient between HOM and DIS is highly related to each other. The correlation coefficient between ASM and ENT and that between ASM and DIS are both larger than 0. 6. DIS has rather strong correlations with © 2013 ACADEMY PUBLISHER
B. Gray Features Extraction The quality of image segmentation is largely dependent on the features extracted from the images. It can be seen from Fig. 6 that it is impossible to get good segmentation only with textural features. That is why we applied gray features more than GLCM features based on NSCT domain. In an SAR image, pixel gray value f (i, j ) is generally regarded as the feature since there are certainly differences in different ground clutter energies. Gray variance D reflects the discreteness of the offcentered pixel gray value in an image and it has the ability of distinguishing different ground features. We regarded the current pixel gray value and surrounding pixel variances as the current pixel gray statistics feature. The properties of local domain were considered compared with the method to only take its gray value as the feature, so this method could be used for representing area attributes with good anti-noise performance. Assuming that f (i, j ) was the pixel gray value in the i th row and j th column, for a window W containing M×N pixels and taking (m, n) as the center, the pixel gray standard deviation D in this domain could be calculated through formula (20):
M N 1 f (i, j ) MN i 1 j 1 1 M N 2 D MN [ f (i, j ) ] i 1 j 1
(20)
We conducted feature quantity fusion over the above extracted gray features and GLCM textural features based on NSCT domain and formed multi-feature vector. Then, there were totally 28 dimension:
V (m, n) [ f (m, n), D(m, n)SN _ G(m, n)]i 1,
,13
(21)
Which represent the features of current pixel V (m, n) . B. Parameter Selection and Optimization of SVM The features extracted are divided into feature spaces by classifier to achieve image segmentation. Since SVM has a high classification accuracy compared with other classifiers, we selected SVM as the classifier and used LIBSVM [17] toolbox to achieve the division of feature vector. The following are the parameter selections for SVM: 1) Kernel function selection We need a kernel function to solve nonlinear classification problems with SVM. The classification performances of different kernel functions vary. Since
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
inner product radial basis function (RDF) has a strong classification capacity and a wide scope of application and it needs few parameters, we selected it as the kernel function of the SVM model. 2) Parameter optimization After selecting RBF kernel function, we need to determine a kernel parameter to maximize the effectiveness of kernel function. The penalty factor of quadratic programming in SVM had a great impact over the classification capability of classifier and we had to select them in a proper way. We optimized the parameter combination through parametric search, in order to confirm a good (C , ) combination and give full play of the classifier performance. Here we applied grid search [18] to seek model parameters. We selected a certain scope of (C , ) , such as =210 ～210 , and assumed that the step size in search was -1; C=210 ～210 , and the step size in search was 1; then we created a two-dimensional grid in (C , ) coordinate system. By cross validation, we calculated the classification accuracy for each set of (C , ) value in corresponding grid and chose the (C , ) value with the highest classification accuracy as the optimal parameter. 3) Selection of training sample Since SVM is acquired by learning from the training sample, the selection of training sample has certain effects over the generalization performance of classifier. For any actual problems, adequate and well distributed training samples lay good foundation for the learning performance of classifier. There are two ways to select the training samples, namely automatic stochastic method and manual selection method. We chose manual selection and selected 50 training samples each for two typical categories to train SVM and then divided the whole image with trained SVM. V.
EXPERIMENTAL RESULTS AND ANALYSIS
We applied simulation SAR images and real SAR to conduct experiment for above-said SAR image segmentation, so as to test the effectiveness of the feature extraction method in the paper. First we used synthetic image (knowing clearly the segmentation result) to imitate SAR images and compared the segmentation results of various algorithms with the known segmentation results (standard segmentation model). Then we evaluated the feature extraction methods by calculating the segmentation accuracy. This was because the actual effectiveness of SAR images was unknown when conducting segmentation and evaluation over the SAR images and there were no standard segmentation samples. We had to regard the manual segmentation results as the reference images and took the segmentation accuracy of various feature extraction as the evaluation index. At the same time, we evaluated with the subjective vision, namely, the intuitive views of the evaluators over the results (visual effect evaluation and empirical evaluation of the evaluators) and decided the quality of the segmentation images by manual vision. For the whole image, comparing the image effects before and after © 2013 ACADEMY PUBLISHER
681
segmentation, we could find out whether the segmentation results were effective. Since it was impossible to evaluate directly the feature data extracted, we applied indirect evaluation. Namely, we extracted SAR image features by different feature extraction methods and divided the feature space through SVM to achieve image segmentation and at last we evaluated the feature extraction through segmentation accuracy. Segmentation accuracy is defined as dividing the total number of pixels of the image by correctly classified pixels of the image, as shown in formula (21). This index can roughly reflect the quality of various segmentation algorithm performances (segmentation accuracy). Segmentation accuracy
Correclty classified pixels (22) Total number of pixels of the image
1) Simulated segmentation experiment of synthetic SAR images We conducted simulation experiments using synthetic SAR image, so as to avoid the errors of manual segmentation when we calculated segmentation accuracy and achieve more objective evaluation index. The synthetic image we chose in the paper included a circle and a square with a size of 256×256, as shown in Fig. 7(a). After adding multiplicative noise whose variance was 0. 1 to it, we got a synthetic image as shown in Fig. 7(b) to simulate actual SAR image. We compared the method proposed with GLCM feature extraction method based on wavelet domain [8] to test the effectiveness of GLCM feature extraction based on NSCT domain. We conducted feature extraction respectively by the two methods and completed image segmentation with SVM; the segmentation results were shown in Fig. 7(c) and Fig. 7 (d). See the second section (SAR image segmentation experiment) for parameters selection of the two feature extraction methods. The contrasted results of accuracy are shown in Table II.
(a) Original synthetic image (b) SAR emulational image added with multiplicative noise
(c) GLCM segmentation image based on wavelet domain (d) GLCM segmentation image based on NSCT domain Figure 7. Synthetic textural image segmentation
682
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
TABLE II.
SEGMENTATION ACCURACY TWO FEATURE EXTRACTION METHODS
Feature extraction GLCM f GLCM method Segmentation 0. 9833 0. 9892 accuracy
Feature extraction GLCM method Segmentation 0. 9833 accuracy
From Fig. 7(c) and Fig. 7(d) it could be seen that the image was segmented into black area and white area. Each area contains several small heterogeneous areas (segmentation fragments). This was because the pixels were segmented incorrectly due to the impacts of noise. The noise points in Fig. 7(d) were less than that in Fig. 7c) and the edge contour was smoother than that of Fig. 7(c). Table II showed that the method proposed has high segmentation accuracy with 0. 059 higher than that of GLCM feature extraction method based on wavelet domain. Combining Fig. 7(c), Fig. 7(d) and Table II, the GLCM feature extraction method based on NSCT domain could effectively resist the impact of noise and provide correct segmentation accuracy and smooth edge comparing with GLCM feature extraction method based on wavelet domain.
segmentation effects. Fig 10 showed the segmentation results of the two feature extraction methods (the white representing the land, and the black representing the water area); Fig. 10(a) showed the segmentation result of GLCM feature extraction based on wavelet domain and Fig. 10(b) showed the segmentation result of GLCM feature extraction method based on NSCT domain, namely the proposed method.
Figure 9. Manual segmentation image
(a) GLSM segmentation image based on wavelet domain (b) GLSM segmentation image based on NSCT domain Figure 8. Original SAR image (River)
2) SAR images segmentation experiment In terms of the parameter selection of GLCM feature extraction algorithm based on NSCT domain, we selected “dmaxflat7” for tower type filter bank and applied twolayer decomposition to get a low-frequency sub-band and 12 high-frequency sub-bands; we selected “maxflat” for direction filter bank and the decomposition level is [2 3]. The feature extracted contained the 26-dimensional textural feature quantity of HOM and ASM and 2dimensional gray feature quantity by GLCM feature extraction method based on NSCT domain. In the GLCM feature extraction algorithm based on wavelet domain for comparison, we selected “db1” for the wavelet basis and applied also two-layer decomposition to get a lowfrequency sub-band and 6 high-frequency sub-bands. The features extracted contained 28-dimensional symbiosis amount of HOM, ASM, ENT and DIS and 2-dimensional gray feature quantity based on wavelet domain. To test the effectiveness of the segmentation method proposed in the paper, we selected a SAR image (river, with a size of 376*375) with abundant detailed information to conduct experiment. To reduce the impact of noise in the SAR image, we conducted noise reduction before feature extraction and the results were shown in Fig. 5. Fig. 8 showed the original SAR image (river) in the experiment and Fig. 9 showed the corresponding manual segmentation image as a reference template for calculating segmentation accuracy and visual
© 2013 ACADEMY PUBLISHER
Figure 10. Segmentation results of the two feature extraction methods
It could be seen from the segmentation results in Fig. (a) and Fig. (b) that the segmentation effect in Fig. 10(b) was better than that in Fig. 10(a). The three details marked with boxes in the two images showed the differences between the two feature extraction methods: Fig. 10(a) showed an obvious deficiency in segmentation; several details were not segmented and there were serious incorrect segmentations; the edge was not smooth enough; some parts segmented were even discontinuous, as shown in the yellow part at the uppermost part of Fig. 10(a), while the edges were continuous in actual SAR image. Relatively, Fig. 10(b) segmented the thin river better and the river details were continuous. As shown in the uppermost yellow area in Fig. 10(b), the thin river branches were well segmented from the background. The image edges were distinct and smooth and the segmentation edge coincided with the original image; the areas segmented presented a good consistency (check whether areas with relatively consistent textural features in the original images presented a unified area). From the segmentation results in the contrast experiments it could be seen that the method in the paper was able to keep detailed information of image edges; it could inspect more boundaries with directivity in SAR images; and it could keep finer details. Meanwhile, we compared the above-mentioned experiment result with the segmentation result and calculated the segmentation accuracy. In the process of accuracy calculation, we compared the algorithm in Fig.
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
10(a), Fig. 10(b) respectively with Fig. 9 and the results were shown in Table III. TABLE III.
SEGMENTATION ACCURACY OF THE TWO FEATURE SEGMENTATION METHODS
Feature extraction GLCM fe GLCM f method Segmentation 0. 9156 0. 9230 accuracy
Feature extraction GLCM method Segmentation 0. 9156 accuracy
From Table III we knew that the segmentation accuracy of feature extraction method proposed in the paper was 0. 074 higher than that of GLCM based on wavelet domain. It showed that the method we used could effectively reduce incorrect segmentation of pixels and it showed indirectly that the feature extraction method proposed was better than GLCM feature extraction method based on wavelet domain. To better compare the segmentation details, we took one of the three places in Fig. 8 (the purple box in lower right corner) as an example to explain the difference of detail segmentation of the two feature extraction methods. Fig. 11(a) to Fig. 11(c) were the enlarged drawing of details corresponding to the purple boxes in Fig. 8, Fig. 10(a) and Fig. 10(b). From Fig. 11 it could be seen that, since SAR image was greatly affected by noise, neither of the two methods could segment the details in the SAR images and there were deficiency in segmentation to some degree (some detailed areas describing the images were not segmented from the background). However, the GLCM feature extraction method based on NSCT domain did better in this than GLCM feature extraction method based on wavelet domain; its segmentation results and visual observation fitted better with the visual interpretation of original SAR images and provided better segmentation effects.
683
Moreover, we compared the experiment results in this paper with segmentation result in literature [19]. Fig. 12 was the single polarization SAR image applying radarsat satellite. Fig. 13(a) was the segmentation result in literature [19] and Fig. 13(b) was the segmentation result in the paper. From Fig. 13(b) it could be seen that edges of the image were distinct and smooth relatively with few segmentation fragments and edges of segmentation result coincided with that of the original image.
Figure 13. Original SAR image (city)
The difference between the proposed method and literature [19] were: I. We conducted noise reduction before feature extraction and reduced the impact of noise over the features; II. The processing unit of the feature extraction was not small image but pixel, which could better reflect the attributes of areas the pixel belonged to; III. We applied GLCM feature extraction method based on NSCT domain to extract the textural features of SAR images and achieve multi-scale and multi-direction textural feature extraction; IV. We optimized the model parameter (C , ) of SVM with grid search method and improved the classification performance of SVM. It could be seen from Fig. 7, Fig. 10, Fig. 13, Table II and Table III that the feature extraction method proposed could reduce incorrect segmentation of pixels and achieve better edge-preserving effects; there was nice goodness of fit between the segmentation images and the original SAR images. VI.
(a) Detailed parts in original images (b) Detailed parts in Fig. 10(a) (c) Detailed parts in Fig. 10(b) Figure 11. Detail comparison of segmentation results of the two feature extraction methods
(a) Segmentation image in literature [19] (b) Segmentation image of method in the paper Figure 12. Segmentation results comparison diagram between the method in the paper and that in literature [19]
© 2013 ACADEMY PUBLISHER
CONCLUSIONS
SAR image segmentation is the basis and premise to interpret the SAR images. However, due to the particularity of SAR images, general image segmentation methods tend to losing effectiveness. Therefore, it is of significant importance to study SAR segmentation algorithm with high accuracy of segmentation. We proposed a GLCM textural feature extraction method based on NSCT domain and a multi-feature vector was formed combining with gray feature; at the same time, through correlation analysis, we selected the feature quantity extracted to eliminate the redundancy of feature quantity and reduce calculation amount and storage space. At last we used the new tool SVM with good performance to divide the feature quality and achieved the segmentation of SAR images. Experimental result showed that the method proposed in the paper produced detailed and smooth boundary direction information, eliminated effectively the incorrect segmentation in the domain and provided accurate segmentation result. By extracting symbiosis amount in NSCT domain, a multi-
684
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
scale and multi-direction textural feature extraction was realized; the shortage that wavelet directivity was limited and that it was not the optimal basis of line singularity was overcome; textural feature could be better described. The experimental results showed the effectiveness of multi-feature fusion. However, it did not mean that the greater the quantity is, the better the features are; it was the features in mutual complementarities that should to be selected. In the next step, we should apply GLCM in other multi-scale transformations, such as directionlet and others; in terms of feature extraction, we should seek for the features that can describe the inherent attributes of objects in images, such as extracting stable textural features with different feature extraction methods; the sample selection in the paper was selected by manual work and to realize automatic selection, we should make use of FCM and achieve automatic selection of samples. ACKNOWLEDGMENT This paper is supported by the provincial-level key discipline of enterprise management in Zhejiang and the construct program of the key laboratory in Hangzhou. REFERENCES [1] An Chengjin, Niu Zhaodong, Li Zhijun et al. Threshold Comparison of Typical Otsu Algorithm and Water Area Segmentation Performance Analysis of its SAR Image. Journal of Electronics Information Technology, 2010, 32(9) pp. 2215-2219. [2] Jiao Licheng. Intelligent SAR Image Processing and Interpretation. Beijing: Science Press, 2008. [3] Kandaswamy U, Adjeroh D A, Lee M C. Efficient texture analysis of SAR image. IEEE Transactions on Geoscience and Remote Sensing, 2005, 43(9) pp. 2075-2083. [4] Wang Zhi-Zhong, Yong Jun-Hai. Texture analysis and classification with linear regression model base on wavelet transform. IEEE Transactions on Image Processing, 2008, 17(8) pp. 1421-1430. [5] Su Hui, Fei Shumin. Fabric Texture Recognition Based on Multi-resolution Analysis and Gray-level Co-occurrence Matrix: the second term of Chinese Control Conference. Beijing: Press of Beijing University of Aeronautics and Astronautics, 2006 pp. 1867-1871. [6] Han Yanfang, Shi Pengfei. Texture Surface Damage Inspection Based on the Multi-layer Wavelets and Cooccurrence Matrix. Journal of Shanghai Jiaotong University, 2006, 40(3) pp. 425-430.
© 2013 ACADEMY PUBLISHER
[7] Shang Zhaowei, Zhang Mingxin, Zhao Ping et al. Texture Retrieval and Similarity Calculation Based on Different Complex Wavelet Transform Methods. Journal of Computer Research and Development, 2005, 42(10) pp. 1746-1751. [8] Zhong Hua, Yang Xiaoming, Jiao Licheng. Textural image classification Based on Multi-resolution Co-occurrence Matrix. Journal of Computer Research and Development, 2011, 48(11) pp. 1991-1999. [9] CLAUSI D A. Improved Texture Recognition of SAR Sea Ice Image by Data Fusion of MRF Features with Traditional Methods. IEEE Transactions on Geoscience and Remote Sensing Symposium, 2001, 3 pp. 1170-1172. [10] Wang Haifeng, Li Zhuang, Ren Honge et al. Textural Image Segmentation Algorithm Based on Non-subsample Contourlet Transform and SVM: the 29th term of Chinese Control Conference Proceedings. Beijing, 2010 pp. 27122716. [11] Wang Qingxiang, Li Di, Zhang Wujie. Non-supervision Segmentation Based on Multi-feature SAR Images. Computer Science, 2010, 37(10) pp. 267-270. [12] CUNHA A L, ZHOU Jian-Ping, DO M N. The Nonsubsampled Contourlet Transform: Theory Design, and Applications. IEEE Transactions on Image Processing, 2006, 15(10) pp. 3089-310l. [13] HARALICK R M, SHANMUGAM K, DINSTEIN L. Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics, 1973, 3(6) pp. 610-621. [14] Bo hua, Ma Fulong, Jiao Licheng. Analysis of Gray Cooccurrence Matrix Computational Problems of Image Texture. Chinese Journal of Electronics, 200634(1) pp. 155-158. [15] FABIO D A, PAOLO G. Texture-based characterization of urban environments on satellite SAR images. IEEE Transactions on Geoscience and Remote Sensing, 2003, 41(1) pp. 153-159. [16] Zhang Xuegong. About Statistics Learning Theory and Support Vector Machine. Acta Automatica Sinica, 2000, 26(1) pp. 32-42. [17] Chang Chih-Chung, Lin Chih-Jen. LIBSVM toolbox [EB/OL]. (2012-11) [2012-11-22]. http: //www. csie. ntu. edu. tw/~cjlin. [18] Wang Xingling, Li Zhanbin. Determination of Kernel Function Parameters of Support Vector Machine Based on Grid Research. Journal of Ocean University of China, 2005, 35(5) pp. 859-862. [19] Huang liang, Shu Ning. Simple Target SAR Image Segmentation Based on Support Vector Machine. Geospatial Information, 2011, 9(1) pp. 60-62.
675
Texture Feature Extraction Method Combining Nonsubsampled Contour Transformation with Gray Level Co-occurrence Matrix He Xiaolan and Wu Yili Business School, Zhejiang University City College, Hangzhou 310015, China
Wu Yiwei Valparaiso University, IN 46383, USA
Abstract—Gray level co-occurrence matrix (GLCM) is an important method to extract the image texture features of synthetic aperture radar (SAR). However, GLCM can only extract the textures under single scale and single direction. A kind of texture feature extraction method combining nonsubsampled contour transformation (NSCT) and GLCM is proposed, so as to achieve the extraction of texture features under multi-scale and multi-direction. We firstly conducted multi-scale and multi-direction decomposition on the SAR images with NSCT, secondly extracted the symbiosis amount with GLCM from the obtained sub-band images, then conducted the correlation analysis for the extracted symbiosis amount to remove the redundant characteristic quantity; and combined it with the gray features to constitute the multi-feature vector. Finally, we made full use of the advantages of the support vector machine in the aspects of small sample database and generalization ability, and completed the division of multifeature vector space by SVM so as to achieve the SAR image segmentation. The results of the experiment showed that the segmentation accuracy rate could be improved and good edge retention effect could be obtained through using the GLCM texture extraction method based on NSCT domain and multi-feature fusion in the SAR image segmentation. Index Terms—Synthetic Aperture Radar; Image Segmentation; Nonsubsampled Contour Transformation; Gray Level Co-Occurrence Matrix; Support Vector Machine; Feature Selection
I.
INTRODUCTION
Synthetic aperture radar (SAR) can form images at all time and all weather, and it plays an increasingly significant role in the military and civilian aspects. Therefore, each country pays great attention to the research and application of the relevant technology of SAR. SAR image interpretation is a very important research field of SAR, while SAR image segmentation is the key pretreatment process in the SAR image interpretation. Because the segmentation quality of the images directly influence the late target recognition of SAR images (Such as airports, ports, roads and bridges), the image segmentation is viewed as the bottleneck in the SAR image interpretation [1]. The general image
© 2013 ACADEMY PUBLISHER doi:10.4304/jmm.8.6.675-684
segmentation methods often fail due to the imaging particularity of SAR images. Therefore, it is very necessary to research effective SAR image segmentation algorithm. At present SAR image segmentation is roughly divided into thresholding method, edge segmentation method, statistical segmentation method and the segmentation method based on pixel features [2]. The thresholding method, edge segmentation method, and statistical segmentation method all use the gray information to achieve the segmentation. However, it is difficult to achieve good SAR image segmentation just by using gray features. The reason is because SAR images reflect the backscattering characteristics of ground objects on the radar wave. If different ground objects have same or similar backscattering coefficients, they will exhibit same or similar gray values in the SAR images, and accordingly confusion will occur. The texture information of SAR images is very rich. The different shapes and different physical features of the imaging targets show different textures, and these textures can become the important basis for the radar to distinguish different objects. The segmentation method based on pixel features can make full use of the intrinsic properties of the images (Such as texture features), and conduct multi-feature fusion through extracting various different features, so as to achieve high segmentation precision. Gray level co-occurrence matrix (GLCM) is the common method to describe the textures. However, its descriptions of the texture features are not so detailed that it is not good enough to extract the features by this method. In order to make full use of the advantages of GLCM and multi-resolution transformation, many scholars [4] have researched the texture feature extraction method of multi-scale GLCM. Su Hui [5] proposed a method to conduct the symbiosis feature extraction on the approximation sub-band on each scale of wavelet transform. This method can describe the texture structure effectively under different resolutions and it is successfully used for the fabric texture recognition, but it ignores to use the detail sub-band of wavelet transform. Han Yanfang [6] proposed to use GLCM in the wavelet
676
detail sub-band to extract the symbiosis features of the texture, and used it for the defect detection of texture surface. But it is insufficient to describe different texture details just with wavelet detail sub-band. Shang Zhaowei [7] et al. extracted 7 symbiosis features of coefficient sub-band in the complex wavelet transform domain and used them for texture retrieval, but no consideration was given to the problems of parameter optimization and feature selection when extracting the symbiosis features. In view of the above problems, Zhong Hua [8] et al. proposed the multi-resolution co-occurrence matrix feature extraction method through extracting the cooccurrence matrix on the approximation sub-band and detail sub-band of the nonsubsampled wavelet transform simultaneously, so as to obtain good classification results in the classification of texture images. Clausi [9] et al. comprehensively used the airspace gray level cooccurrence matrix and Gabor wavelet to extract the texture features of SAR images, and thus conducted SAR sea ice image segmentation. The experiments showed that the classification accuracy of using different ways to extract the texture feature for fusion was higher than that of the classification only using a kind of texture. Wang Haifeng [10] et al. proposed a kind of texture image segmentation algorithm combining nonsubsampled contour transformation (NSCT) and support vector machine (SVM), used NSCT to decompose the images, then extracted the texture features, and finally completed the segmentation of the entire feature image by SVM. However, only the multi-scale feature of the image is extracted through this method, while the spatial structure information of the texture is ignored. In view of the features such as large gray change, complex texture and fuzzy boundaries, Wang Qingxiang [11] et al. proposed a kind of unsupervised segmentation method of SAR images based on multi-features, and obtained good segmentation effect. The several feature extraction methods mentioned above are all conducted in the airspace or wavelet domain. However, the wavelet transform has limited isotropy and directionality (Three), there is no translation invariant, and the singularity on the line is not the optimal basis, so the texture information is not fully shown. As a kind of multi-scale geometric analysis tool, NSCT has the optimal representation form on the curve singularity function, it can overcome the above-mentioned shortcomings of the wavelet transform, and it has obtained good effect in the field of the image analysis (such as noise reduction, feature extraction and retrieval). GLCM is a kind of texture analysis method which is used most widely and it is commonly used to extract the texture features of the images. The multi-resolution and multi-direction characteristics of NSCT can be combined with the texture structure information of GLCM to extract the texture features of SAR images with GLCM in the NSCT domain, so that the gray symbiosis features can provide the dynamic information of texture space scale change and have better descriptive ability for the texture. At the same time, we can give full play to the characteristics of multi-direction of NSCT by extracting
© 2013 ACADEMY PUBLISHER
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
the symbiosis amount of different sub-band in the NSCT domain (Each sub-band of NSCT represents different direction information.). In this paper, we have researched the GLCM texture feature extraction method based on NSCT domain. The major innovation points are as follows: 1. The anisotropy, multi-resolution and multidirection characteristics of NSCT are used, multiscale and multi-direction decompositions are conducted for the SAR images by using NSCT, the sub-band images of different scales and different directions are obtained, the high-dimension singularity information and rich texture information in the images are captured, and the limitations of the wavelet transform are overcome. 2. The spatial distribution information and structure information property of the images can be described using GLCM. NSCT is combined with GLCM, and GLCM is used for the sub-band images of different scales and different directions after passing NSCT to extract the gray symbiosis feature quantity, so as to overcome the limitations that wavelet domain GLCM can only extract the gray symbiosis amount in the limited directions and it cannot describe the line singularity effectively, and thus achieve the extraction of the texture features of multi-resolution and multi-direction. 3. In this method, the extracted gray values and gray variances are taken as the gray features, the multifeature fusion strategy (namely combining texture feature and gray features) is used, so as to constitute the multi-feature values and use them for the SAR image segmentation. 4. Since the support vector machine (SVM) has higher classification accuracy than other classifiers in the aspects such as high-dimension feature vector and small sample data, in this method SVM is used to complete the division of high-dimension feature space so as to achieve the purpose of segmenting SAR images. The experimental results show that the GLCM feature extraction method based on NSCT domain is an effective method to extract the features of SAR images and its segmentation accuracy rate exceeds that of GLCM feature extraction method of wavelet domain. This method can improve the accuracy of the segmentation image edge location and reduce the misclassification of the pixels within the region, so as to obtain good segmentation results. II.
GLCM FEATURE EXTRACTION METHOD BASED ON NSCT DOMAIN
A. Nonsubsampled Contour Transformation Nonsubsampled Contour Transformation [12] (NSCT) is a kind of algorithm which uses the iterative nonsubsampled filter bank to achieve a series of multiresolution, multi-direction and translation invariant frequency domain sub-images. In this algorithm, multiscale analysis is conducted firstly on the images, then multi-direction analysis is conducted, and the layer by
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
layer iteration is conducted separately for these two steps. The structure of NSCT is divided into nonsubsampled pyramid filter bank (NSPFB) and nonsubsampled directional filter bank (NSDFB), and its realization structural principle is as shown in Fig. 1. Its decomposition process can be expressed as: Firstly the nonsubsampled pyramid filter bank (NSPFB) is used to conduct multi-scale decomposition on the images, so as to obtain various high-frequency sub-band images with different frequencies and a low-frequency sub-band image, and then the nonsubsampled directional filter bank is used to conduct multi-direction decomposition for the obtained various high-frequency sub-band images, so as to obtain the sub-band images (coefficients) of different scales and different directions. Low pass subband Second stage low pass filter
High-frequency directional sub-band
Image
677
pixels in y direction is Ny, and the highest gray level of the images is Ng level.
Figure 2. Schematic diagram of gray level co-occurrence matrix structure
Starting from the pixel with a gray value of i , the gray value occurring simultaneously in another pixel with a straight-line distance of d and an angle of is j ; based on this, the probability of these two gray values occurring simultaneously in the entire images can be defined as: P(i, j, ) {( x, y) | f ( x, y) i, f ( x d cos , y d sin ) } (1)
High-frequency First stage low pass filter directional sub-band
Nonsubsampled pyramid Nonsubsampled directional filter bank filter bank
Figure 1. NSCT realization structural principle diagram
When the decomposition level of NSPFB is L, 1 low pass sub-band image and L high pass sub-band images can be obtained through decomposition. Because there is no down-sampling, all the sub-band images have the same sizes with the original images. The realization process of NSDFB is that firstly up-sampling is conducted on the filter, and secondly the wave filtering is conducted on the decomposed sub-band images in the two channel direction of last stage, so as to achieve the more accurate directional decomposition in the frequency domain. 2t directional sub-band images which have the same size as the original images can be obtained through l stage directional decomposition of NSDFB on a scale sub-band image. NSPFB is combined with NSDFB to achieve NSCT. The band-pass sub-band images, which are obtained after decomposing the original images through NSPFB, are inputted NSDFB and the band-pass directional sub-band images can be obtained, so as to achieve the multi-scale and multi-direction decomposition of images. Gray level co-occurrence matrix (GLCM) [13] is recognized as an important texture analytical method today. Its definition is as follows: In Fig. 2, oxy is set as the coordinate plane of the image pixel, the gray coordinate is z axis, the total number of pixels in x direction is Nx, the total number of
© 2013 ACADEMY PUBLISHER
The probability of two pixel gray levels occurring simultaneously converts the spatial coordinates of (x, y) to the gray descriptions of (i, j) and forms the gray level co-occurrence matrix. In the actual texture analysis, the gray level co-occurrence matrix often is not used directly, while the second degree statistics are extracted on the basis of this matrix and used as the feature values of texture identification. A large number of experiments showed that for the SAR images the following four statistics could best express the texture features of SAR images [14][15]. Homogeneous area:
HOM p(i, j ) / [1 (i j )2 ] i
(2)
j
Angular second moment:
ASM [ p(i, j )]2
(3)
ENT p(i, j ) log( p(i, j ))
(4)
i
j
Entropy: i
j
Dissimilarity:
DIS | i j | p(i, j ) i
III.
(5)
j
PRINCIPLES OF SUPPORT VECTOR MACHINE
The principle of the support vector machine (SVM) is that we use the classification hyper plane H: w, x b 0 to correctly separate the two sample points {( x1 Rn , yi {1, 1}, i 1, 2,.., N ) in the space, H1:
w, x b 1 , H2: w, x b 1 , as shown in Fig. 3, and we obtain the minimum distances of the positive
678
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
sample and negative sample to the hyper plane, so as to make the interval between the two samples M arg in 2 / w H2 H1
H
the high-dimension feature space can be obtained as follows:
f ( x) sgn(k ( w, x) b) sgn yi ai K ( x, xi ) b iSV
Support vector
(9)
In the practical application, the commonly used kernel functions are as follows: Linear kernel function (linear):
K ( xi , x j ) xi x j
Margin=2/ w
(10)
Radial basis kernel function (RBF): Figure 3. Schematic diagram of optimal classification hyper plane
The original problem is mathematically described as a constrained nonlinear programming problem: 1 2 Minimize ( w, b) w 2
yi ( xi w b) 1 0 i 1, 2,
,N
(6)
The minimum norm W meeting the constraint is the normal vector of the optimal classification hyper plane. The decision function based on the optimal classification hyper plane is as follows:
f ( x) sgn yi ai x, xi b iSV
(7)
In the formula, SV is the support vector, and ai 0 is the unique solution of the above-mentioned secondary programming problem. As for the nonlinear case, we just need to replace the dot product in the dual problem with the convolution kernel function K ( xi , x j ) . As for the case that the sample points cannot be separated, we can introduce the slack variable i 0(i 1, , N ) and penalty factor C 0 , where, C is used for controlling the punishment degree for the misclassified samples. The soft edge classification surface is constructed so as to allow the misclassified samples to exist. The mathematical model of the corresponding optimization problem is as follows: N 1 2 Minimize ( w, b) w C i 2 i 1
yi ( xi w b) 1 i i 0, i 1, 2,
,N
(8)
From the statistical learning theory we can obtain that, the function K ( x, y) meeting the Mercer Theorem can be used as the inner product, and the nucleation function
K ( x, y ) i i ( x) i ( y) can be expanded with the i 1
coefficient i 0 . After the kernel function is introduced, when the sample is not linearly separable in the highdimension feature space, the decision function of the generalized optimal classification hyper plane based on
© 2013 ACADEMY PUBLISHER
2
K ( xi , x j ) exp{ xi x j }
(11)
Polynomial kernel function (polynomial):
K ( xi , x j ) [( xi x j ) ]d }
(12)
Two-layer neural network kernel function (Sigmoid tanh):
K ( xi , x j ) thah(u( xi x j ) c) IV.
(13)
EXPERIMENT PROCESS
A. GLCM Texture Feature Extraction and Selection Based on NSCT Domain Assuming that the input image is f ( x, y ) and the size is M*N, the decomposition process of NSCT can be expressed as: J
lj
f ( x, y ) NSCTaJ b j , k
(14)
j 1 k 1
where, aJ is the low-frequency sub-band, and b j , k is the high-frequency sub-band of scale j and direction k. A NSCT transform coefficient can be expressed with b( j, k , m, n) , where, J is the decomposition level of NSPFB, is the decomposition level of NSDFB of layer j; j represents the scale label after the decomposition of NSPFB, k represents the direction label in NSDFB, and m and n represent the spatial location information of the coefficients in the directional sub-bands. GLCM is applied to decompose NSCT so as to obtain the lowfrequency sub-band aJ and the high-frequency sub-band
b j , k , the homogeneous area (HOM), angular second moment (ASM), entropy (ENT) and dissimilarity (DIS) of the gray symbiosis amount are extracted, the extracted symbiosis amounts are taken as the texture feature vectors, and thereby the GLCM texture feature extraction based on NSCT domain is completed. The extraction steps of the GLCM texture feature quantity based on NSCT domain are as follows: 1) NSCT decomposition of images. Firstly we use NSCT to conduct two-layer decomposition for the images. The first scale direction number is 8, the second scale direction number is 4, and
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
13 sub-bands with the same size as the original images are obtained. In this way, the images are decomposed into the sub-bands with different scales and different directions, and each directional sub-band represents a direction of texture. 2) Transform coefficient quantification. NSCT low-frequency sub-band approximately obeys the uniform distribution, while the high-frequency subband shows a distribution of “High Peak, Long Tail”, as shown in Fig. 4 (The horizontal axis represents the amplitude, and the vertical axis represents the frequency.). The NSCT coefficients and wavelet coefficients have similar statistical distribution, and thus the coefficients are quantified according to the quantification strategy of the literature [8]. The uniform quantification is applied to the low-frequency sub-bands and the quantification level is 16, that is to say, the quantification is conducted according to the formula (15), where, [] represents rounding down.
G [(C Cmin ) / (Cmax Cmin )] 16
(15)
The non-uniform quantification is applied to the detail sub-band, and the coefficient is quantified to 16 levels. Each sub-band variance l of the same scale is estimated firstly, is taken as the boundary during the quantification, and then the piecewise quantification is conducted according to formula (16).
Gl
[( C Cmin )/( Cmax Cmin )]12,|C ( m , n ) 3 l [( C Cmin )/( Cmax Cmin )] 2,|C ( m, n ) 3 l
(16)
679
we applied windows of N=5×5 for sub-band images in second scale and low-frequency sub-bands. 4) Distance parameter d. Distance d reflects the frequency period of texture. The number of direction change varies in accordance with different SNCT decomposition parameter settings, and there is certain difference between different dimensions with the change of direction parameters. Therefore, different distance intervals d should be applied for different scales when calculating the gray symbiosis amount for NSCT decomposition sub-band. A small distance is applied for rough texture (high-frequency) and a large distance is applied for delicate texture (low-frequency sub-band). According to this, we applied d=1 for changing 8 subbands in the first scale and applied d=2 for changing 4 sub-bands in the second scale and low-frequency subbands. 5) Selection of direction . We drew the symbiosis amounts through GLCM formula (2) to (5) and calculated the homogeneous area, angular second moment, entropy and dissimilarity in four angles ( were respectively 0 , 45 , 90 and 135 ) of the NSCT sub-bands, so as to prepare for the following multi-feature fusion. Then we took the average of values in four directions of each symbiosis amount, as shown in formula (17), so as to improve the robustness of gray symbiosis over directions.
f ( f o f 45 f 90 f 135 ) / 4
At last, we got the gray symbiotic feature quantity of the image at point (m, n) in NSCT domain according to the parameter selection principles. We described the textural features of this point with 52-dimensional symbiosis amount:
N _ G(m, n) [ HOM i , AMSi , ENTi , DISi ]i 1,
Figure 4. NSCT coefficient gray statistical histogram
3) Window traversal method and window size are considered when selecting window. In the paper, we applied overlaid windows for window traversal images, that is, we get a small window, by regarding each pixel as the center, the characteristics calculated in the window are only the features of the center point, and the final categorical attributes are only those of the center point. The selection of window size is also critical, since a too large window will lead to large calculation amount and memory space while a too small window can’t contain complete texture information. Theoretically, large windows are applied for low-frequency rough texture; small windows are applied for high-frequency delicate texture. According to the criteria for window size selection, we applied windows of different sizes for different NSCT decomposition scales. We applied windows of N=3×3 for sub-band images in the first scale;
,13
(18)
However, there is certain correlation between the four gray symbiosis amounts as homogeneous area, angular second moment, entropy and dissimilarity. We first conducted correlation analysis over the feature quantity extracted in terms of features selection, and then we eliminated redundant features to reduce the dimensions and calculation amount. Fig. 5 shows the SAR image after noise reduction. We drew the textural features of SAR image in the same way as GLCM feature extraction based on NSCT domain, and then we divided the four textural features as homogeneous area, angular second moment, entropy and dissimilarity by SVM and got the single-feature segmentation image as shown in Fig. 6(a) to Fig. 6(d).
Figure 5. SAR image after noise reduction © 2013 ACADEMY PUBLISHER
(17)
680
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
HOM, ASM and ENT, and the correlation coefficients are all larger than 0. 6. From Table I and Fig. 6(a) to 6(b) we got that we could reduce the feature dimensions from the original 52 dimensions to 26 dimensions by just selecting two statistics amounts of HOM and ASM, SN _ G(m, n) [HOM i (m ,n ), AMSi (m ,n )],i 1, ,13 namely. (a) ASM
(b) HOM
(c) ENT
(d) DIS
Figure 6. Segmentation images based on single textural features
It can be seen from Fig. 6(a) to (d) that Fig. 6(a), Fig. 6(c) and Fig. 6(d) are very similar, i. e. , there are strong correlations between the three feature quantities (ASM, ENT and DIS). Due to certain correlations between HOM, ASM, ENT, DIS, if we conduct feature selection, we should eliminate feature quantities with strong correlation from the original feature quantities and keep those complementary ones contributive to following segmentation, so as to eliminate redundant features, reduce calculation amount and improve segmentation accuracy. Formula (19) is used to calculate the correlation between two images:
r
( A
mn
m
A)( Bmn B )
n
( ( Amn A) 2 )( ( Bmn B ) 2 ) m
n
m
(19)
n
where, A mean2( A) ; B mean2( B) ; mean() means to take the mean value of the matrix. We calculated the absolute values of correlation coefficient between any two of the above four symbiosis statistics amounts in the 13 sub-bands respectively corresponding to the same scale and the same direction. Then we took the average of the 13 sub-bands results and got the correlation coefficient table (Table I). TABLE I.
CORRELATION COEFFICIENT BETWEEN STATISTICS AMOUNTS
HOM ASM ENT DIS
HOM 1 0. 4796 0. 6282 0. 9019
ASM 0. 4796 1 0. 8649 0. 6184
ENT 0. 6282 0. 8649 1 0. 7713
DIS 0. 9019 0. 6184 0. 7713 1
It can be seen from Table I that the correlation coefficient between statistics amounts HOM and ASM is smaller than 0. 5; the correlation coefficient between HOM and ENT and the correlation coefficient between HOM and DIS are both larger than 0. 6, while the correlation coefficient between HOM and DIS is highly related to each other. The correlation coefficient between ASM and ENT and that between ASM and DIS are both larger than 0. 6. DIS has rather strong correlations with © 2013 ACADEMY PUBLISHER
B. Gray Features Extraction The quality of image segmentation is largely dependent on the features extracted from the images. It can be seen from Fig. 6 that it is impossible to get good segmentation only with textural features. That is why we applied gray features more than GLCM features based on NSCT domain. In an SAR image, pixel gray value f (i, j ) is generally regarded as the feature since there are certainly differences in different ground clutter energies. Gray variance D reflects the discreteness of the offcentered pixel gray value in an image and it has the ability of distinguishing different ground features. We regarded the current pixel gray value and surrounding pixel variances as the current pixel gray statistics feature. The properties of local domain were considered compared with the method to only take its gray value as the feature, so this method could be used for representing area attributes with good anti-noise performance. Assuming that f (i, j ) was the pixel gray value in the i th row and j th column, for a window W containing M×N pixels and taking (m, n) as the center, the pixel gray standard deviation D in this domain could be calculated through formula (20):
M N 1 f (i, j ) MN i 1 j 1 1 M N 2 D MN [ f (i, j ) ] i 1 j 1
(20)
We conducted feature quantity fusion over the above extracted gray features and GLCM textural features based on NSCT domain and formed multi-feature vector. Then, there were totally 28 dimension:
V (m, n) [ f (m, n), D(m, n)SN _ G(m, n)]i 1,
,13
(21)
Which represent the features of current pixel V (m, n) . B. Parameter Selection and Optimization of SVM The features extracted are divided into feature spaces by classifier to achieve image segmentation. Since SVM has a high classification accuracy compared with other classifiers, we selected SVM as the classifier and used LIBSVM [17] toolbox to achieve the division of feature vector. The following are the parameter selections for SVM: 1) Kernel function selection We need a kernel function to solve nonlinear classification problems with SVM. The classification performances of different kernel functions vary. Since
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
inner product radial basis function (RDF) has a strong classification capacity and a wide scope of application and it needs few parameters, we selected it as the kernel function of the SVM model. 2) Parameter optimization After selecting RBF kernel function, we need to determine a kernel parameter to maximize the effectiveness of kernel function. The penalty factor of quadratic programming in SVM had a great impact over the classification capability of classifier and we had to select them in a proper way. We optimized the parameter combination through parametric search, in order to confirm a good (C , ) combination and give full play of the classifier performance. Here we applied grid search [18] to seek model parameters. We selected a certain scope of (C , ) , such as =210 ～210 , and assumed that the step size in search was -1; C=210 ～210 , and the step size in search was 1; then we created a two-dimensional grid in (C , ) coordinate system. By cross validation, we calculated the classification accuracy for each set of (C , ) value in corresponding grid and chose the (C , ) value with the highest classification accuracy as the optimal parameter. 3) Selection of training sample Since SVM is acquired by learning from the training sample, the selection of training sample has certain effects over the generalization performance of classifier. For any actual problems, adequate and well distributed training samples lay good foundation for the learning performance of classifier. There are two ways to select the training samples, namely automatic stochastic method and manual selection method. We chose manual selection and selected 50 training samples each for two typical categories to train SVM and then divided the whole image with trained SVM. V.
EXPERIMENTAL RESULTS AND ANALYSIS
We applied simulation SAR images and real SAR to conduct experiment for above-said SAR image segmentation, so as to test the effectiveness of the feature extraction method in the paper. First we used synthetic image (knowing clearly the segmentation result) to imitate SAR images and compared the segmentation results of various algorithms with the known segmentation results (standard segmentation model). Then we evaluated the feature extraction methods by calculating the segmentation accuracy. This was because the actual effectiveness of SAR images was unknown when conducting segmentation and evaluation over the SAR images and there were no standard segmentation samples. We had to regard the manual segmentation results as the reference images and took the segmentation accuracy of various feature extraction as the evaluation index. At the same time, we evaluated with the subjective vision, namely, the intuitive views of the evaluators over the results (visual effect evaluation and empirical evaluation of the evaluators) and decided the quality of the segmentation images by manual vision. For the whole image, comparing the image effects before and after © 2013 ACADEMY PUBLISHER
681
segmentation, we could find out whether the segmentation results were effective. Since it was impossible to evaluate directly the feature data extracted, we applied indirect evaluation. Namely, we extracted SAR image features by different feature extraction methods and divided the feature space through SVM to achieve image segmentation and at last we evaluated the feature extraction through segmentation accuracy. Segmentation accuracy is defined as dividing the total number of pixels of the image by correctly classified pixels of the image, as shown in formula (21). This index can roughly reflect the quality of various segmentation algorithm performances (segmentation accuracy). Segmentation accuracy
Correclty classified pixels (22) Total number of pixels of the image
1) Simulated segmentation experiment of synthetic SAR images We conducted simulation experiments using synthetic SAR image, so as to avoid the errors of manual segmentation when we calculated segmentation accuracy and achieve more objective evaluation index. The synthetic image we chose in the paper included a circle and a square with a size of 256×256, as shown in Fig. 7(a). After adding multiplicative noise whose variance was 0. 1 to it, we got a synthetic image as shown in Fig. 7(b) to simulate actual SAR image. We compared the method proposed with GLCM feature extraction method based on wavelet domain [8] to test the effectiveness of GLCM feature extraction based on NSCT domain. We conducted feature extraction respectively by the two methods and completed image segmentation with SVM; the segmentation results were shown in Fig. 7(c) and Fig. 7 (d). See the second section (SAR image segmentation experiment) for parameters selection of the two feature extraction methods. The contrasted results of accuracy are shown in Table II.
(a) Original synthetic image (b) SAR emulational image added with multiplicative noise
(c) GLCM segmentation image based on wavelet domain (d) GLCM segmentation image based on NSCT domain Figure 7. Synthetic textural image segmentation
682
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
TABLE II.
SEGMENTATION ACCURACY TWO FEATURE EXTRACTION METHODS
Feature extraction GLCM f GLCM method Segmentation 0. 9833 0. 9892 accuracy
Feature extraction GLCM method Segmentation 0. 9833 accuracy
From Fig. 7(c) and Fig. 7(d) it could be seen that the image was segmented into black area and white area. Each area contains several small heterogeneous areas (segmentation fragments). This was because the pixels were segmented incorrectly due to the impacts of noise. The noise points in Fig. 7(d) were less than that in Fig. 7c) and the edge contour was smoother than that of Fig. 7(c). Table II showed that the method proposed has high segmentation accuracy with 0. 059 higher than that of GLCM feature extraction method based on wavelet domain. Combining Fig. 7(c), Fig. 7(d) and Table II, the GLCM feature extraction method based on NSCT domain could effectively resist the impact of noise and provide correct segmentation accuracy and smooth edge comparing with GLCM feature extraction method based on wavelet domain.
segmentation effects. Fig 10 showed the segmentation results of the two feature extraction methods (the white representing the land, and the black representing the water area); Fig. 10(a) showed the segmentation result of GLCM feature extraction based on wavelet domain and Fig. 10(b) showed the segmentation result of GLCM feature extraction method based on NSCT domain, namely the proposed method.
Figure 9. Manual segmentation image
(a) GLSM segmentation image based on wavelet domain (b) GLSM segmentation image based on NSCT domain Figure 8. Original SAR image (River)
2) SAR images segmentation experiment In terms of the parameter selection of GLCM feature extraction algorithm based on NSCT domain, we selected “dmaxflat7” for tower type filter bank and applied twolayer decomposition to get a low-frequency sub-band and 12 high-frequency sub-bands; we selected “maxflat” for direction filter bank and the decomposition level is [2 3]. The feature extracted contained the 26-dimensional textural feature quantity of HOM and ASM and 2dimensional gray feature quantity by GLCM feature extraction method based on NSCT domain. In the GLCM feature extraction algorithm based on wavelet domain for comparison, we selected “db1” for the wavelet basis and applied also two-layer decomposition to get a lowfrequency sub-band and 6 high-frequency sub-bands. The features extracted contained 28-dimensional symbiosis amount of HOM, ASM, ENT and DIS and 2-dimensional gray feature quantity based on wavelet domain. To test the effectiveness of the segmentation method proposed in the paper, we selected a SAR image (river, with a size of 376*375) with abundant detailed information to conduct experiment. To reduce the impact of noise in the SAR image, we conducted noise reduction before feature extraction and the results were shown in Fig. 5. Fig. 8 showed the original SAR image (river) in the experiment and Fig. 9 showed the corresponding manual segmentation image as a reference template for calculating segmentation accuracy and visual
© 2013 ACADEMY PUBLISHER
Figure 10. Segmentation results of the two feature extraction methods
It could be seen from the segmentation results in Fig. (a) and Fig. (b) that the segmentation effect in Fig. 10(b) was better than that in Fig. 10(a). The three details marked with boxes in the two images showed the differences between the two feature extraction methods: Fig. 10(a) showed an obvious deficiency in segmentation; several details were not segmented and there were serious incorrect segmentations; the edge was not smooth enough; some parts segmented were even discontinuous, as shown in the yellow part at the uppermost part of Fig. 10(a), while the edges were continuous in actual SAR image. Relatively, Fig. 10(b) segmented the thin river better and the river details were continuous. As shown in the uppermost yellow area in Fig. 10(b), the thin river branches were well segmented from the background. The image edges were distinct and smooth and the segmentation edge coincided with the original image; the areas segmented presented a good consistency (check whether areas with relatively consistent textural features in the original images presented a unified area). From the segmentation results in the contrast experiments it could be seen that the method in the paper was able to keep detailed information of image edges; it could inspect more boundaries with directivity in SAR images; and it could keep finer details. Meanwhile, we compared the above-mentioned experiment result with the segmentation result and calculated the segmentation accuracy. In the process of accuracy calculation, we compared the algorithm in Fig.
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
10(a), Fig. 10(b) respectively with Fig. 9 and the results were shown in Table III. TABLE III.
SEGMENTATION ACCURACY OF THE TWO FEATURE SEGMENTATION METHODS
Feature extraction GLCM fe GLCM f method Segmentation 0. 9156 0. 9230 accuracy
Feature extraction GLCM method Segmentation 0. 9156 accuracy
From Table III we knew that the segmentation accuracy of feature extraction method proposed in the paper was 0. 074 higher than that of GLCM based on wavelet domain. It showed that the method we used could effectively reduce incorrect segmentation of pixels and it showed indirectly that the feature extraction method proposed was better than GLCM feature extraction method based on wavelet domain. To better compare the segmentation details, we took one of the three places in Fig. 8 (the purple box in lower right corner) as an example to explain the difference of detail segmentation of the two feature extraction methods. Fig. 11(a) to Fig. 11(c) were the enlarged drawing of details corresponding to the purple boxes in Fig. 8, Fig. 10(a) and Fig. 10(b). From Fig. 11 it could be seen that, since SAR image was greatly affected by noise, neither of the two methods could segment the details in the SAR images and there were deficiency in segmentation to some degree (some detailed areas describing the images were not segmented from the background). However, the GLCM feature extraction method based on NSCT domain did better in this than GLCM feature extraction method based on wavelet domain; its segmentation results and visual observation fitted better with the visual interpretation of original SAR images and provided better segmentation effects.
683
Moreover, we compared the experiment results in this paper with segmentation result in literature [19]. Fig. 12 was the single polarization SAR image applying radarsat satellite. Fig. 13(a) was the segmentation result in literature [19] and Fig. 13(b) was the segmentation result in the paper. From Fig. 13(b) it could be seen that edges of the image were distinct and smooth relatively with few segmentation fragments and edges of segmentation result coincided with that of the original image.
Figure 13. Original SAR image (city)
The difference between the proposed method and literature [19] were: I. We conducted noise reduction before feature extraction and reduced the impact of noise over the features; II. The processing unit of the feature extraction was not small image but pixel, which could better reflect the attributes of areas the pixel belonged to; III. We applied GLCM feature extraction method based on NSCT domain to extract the textural features of SAR images and achieve multi-scale and multi-direction textural feature extraction; IV. We optimized the model parameter (C , ) of SVM with grid search method and improved the classification performance of SVM. It could be seen from Fig. 7, Fig. 10, Fig. 13, Table II and Table III that the feature extraction method proposed could reduce incorrect segmentation of pixels and achieve better edge-preserving effects; there was nice goodness of fit between the segmentation images and the original SAR images. VI.
(a) Detailed parts in original images (b) Detailed parts in Fig. 10(a) (c) Detailed parts in Fig. 10(b) Figure 11. Detail comparison of segmentation results of the two feature extraction methods
(a) Segmentation image in literature [19] (b) Segmentation image of method in the paper Figure 12. Segmentation results comparison diagram between the method in the paper and that in literature [19]
© 2013 ACADEMY PUBLISHER
CONCLUSIONS
SAR image segmentation is the basis and premise to interpret the SAR images. However, due to the particularity of SAR images, general image segmentation methods tend to losing effectiveness. Therefore, it is of significant importance to study SAR segmentation algorithm with high accuracy of segmentation. We proposed a GLCM textural feature extraction method based on NSCT domain and a multi-feature vector was formed combining with gray feature; at the same time, through correlation analysis, we selected the feature quantity extracted to eliminate the redundancy of feature quantity and reduce calculation amount and storage space. At last we used the new tool SVM with good performance to divide the feature quality and achieved the segmentation of SAR images. Experimental result showed that the method proposed in the paper produced detailed and smooth boundary direction information, eliminated effectively the incorrect segmentation in the domain and provided accurate segmentation result. By extracting symbiosis amount in NSCT domain, a multi-
684
JOURNAL OF MULTIMEDIA, VOL. 8, NO. 6, DECEMBER 2013
scale and multi-direction textural feature extraction was realized; the shortage that wavelet directivity was limited and that it was not the optimal basis of line singularity was overcome; textural feature could be better described. The experimental results showed the effectiveness of multi-feature fusion. However, it did not mean that the greater the quantity is, the better the features are; it was the features in mutual complementarities that should to be selected. In the next step, we should apply GLCM in other multi-scale transformations, such as directionlet and others; in terms of feature extraction, we should seek for the features that can describe the inherent attributes of objects in images, such as extracting stable textural features with different feature extraction methods; the sample selection in the paper was selected by manual work and to realize automatic selection, we should make use of FCM and achieve automatic selection of samples. ACKNOWLEDGMENT This paper is supported by the provincial-level key discipline of enterprise management in Zhejiang and the construct program of the key laboratory in Hangzhou. REFERENCES [1] An Chengjin, Niu Zhaodong, Li Zhijun et al. Threshold Comparison of Typical Otsu Algorithm and Water Area Segmentation Performance Analysis of its SAR Image. Journal of Electronics Information Technology, 2010, 32(9) pp. 2215-2219. [2] Jiao Licheng. Intelligent SAR Image Processing and Interpretation. Beijing: Science Press, 2008. [3] Kandaswamy U, Adjeroh D A, Lee M C. Efficient texture analysis of SAR image. IEEE Transactions on Geoscience and Remote Sensing, 2005, 43(9) pp. 2075-2083. [4] Wang Zhi-Zhong, Yong Jun-Hai. Texture analysis and classification with linear regression model base on wavelet transform. IEEE Transactions on Image Processing, 2008, 17(8) pp. 1421-1430. [5] Su Hui, Fei Shumin. Fabric Texture Recognition Based on Multi-resolution Analysis and Gray-level Co-occurrence Matrix: the second term of Chinese Control Conference. Beijing: Press of Beijing University of Aeronautics and Astronautics, 2006 pp. 1867-1871. [6] Han Yanfang, Shi Pengfei. Texture Surface Damage Inspection Based on the Multi-layer Wavelets and Cooccurrence Matrix. Journal of Shanghai Jiaotong University, 2006, 40(3) pp. 425-430.
© 2013 ACADEMY PUBLISHER
[7] Shang Zhaowei, Zhang Mingxin, Zhao Ping et al. Texture Retrieval and Similarity Calculation Based on Different Complex Wavelet Transform Methods. Journal of Computer Research and Development, 2005, 42(10) pp. 1746-1751. [8] Zhong Hua, Yang Xiaoming, Jiao Licheng. Textural image classification Based on Multi-resolution Co-occurrence Matrix. Journal of Computer Research and Development, 2011, 48(11) pp. 1991-1999. [9] CLAUSI D A. Improved Texture Recognition of SAR Sea Ice Image by Data Fusion of MRF Features with Traditional Methods. IEEE Transactions on Geoscience and Remote Sensing Symposium, 2001, 3 pp. 1170-1172. [10] Wang Haifeng, Li Zhuang, Ren Honge et al. Textural Image Segmentation Algorithm Based on Non-subsample Contourlet Transform and SVM: the 29th term of Chinese Control Conference Proceedings. Beijing, 2010 pp. 27122716. [11] Wang Qingxiang, Li Di, Zhang Wujie. Non-supervision Segmentation Based on Multi-feature SAR Images. Computer Science, 2010, 37(10) pp. 267-270. [12] CUNHA A L, ZHOU Jian-Ping, DO M N. The Nonsubsampled Contourlet Transform: Theory Design, and Applications. IEEE Transactions on Image Processing, 2006, 15(10) pp. 3089-310l. [13] HARALICK R M, SHANMUGAM K, DINSTEIN L. Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics, 1973, 3(6) pp. 610-621. [14] Bo hua, Ma Fulong, Jiao Licheng. Analysis of Gray Cooccurrence Matrix Computational Problems of Image Texture. Chinese Journal of Electronics, 200634(1) pp. 155-158. [15] FABIO D A, PAOLO G. Texture-based characterization of urban environments on satellite SAR images. IEEE Transactions on Geoscience and Remote Sensing, 2003, 41(1) pp. 153-159. [16] Zhang Xuegong. About Statistics Learning Theory and Support Vector Machine. Acta Automatica Sinica, 2000, 26(1) pp. 32-42. [17] Chang Chih-Chung, Lin Chih-Jen. LIBSVM toolbox [EB/OL]. (2012-11) [2012-11-22]. http: //www. csie. ntu. edu. tw/~cjlin. [18] Wang Xingling, Li Zhanbin. Determination of Kernel Function Parameters of Support Vector Machine Based on Grid Research. Journal of Ocean University of China, 2005, 35(5) pp. 859-862. [19] Huang liang, Shu Ning. Simple Target SAR Image Segmentation Based on Support Vector Machine. Geospatial Information, 2011, 9(1) pp. 60-62.
