CN107564008B

CN107564008B - Rapid SAR image segmentation method based on key pixel fuzzy clustering

Info

Publication number: CN107564008B
Application number: CN201710652334.0A
Authority: CN
Inventors: 尚荣华; 袁一璟; 焦李成; 刘芳; 马文萍; 王蓉芳; 侯彪; 王爽; 刘红英
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2017-08-02
Filing date: 2017-08-02
Publication date: 2020-11-24
Anticipated expiration: 2037-08-02
Also published as: CN107564008A

Abstract

The invention provides a rapid SAR image segmentation method based on key pixel fuzzy clustering, which is used for solving the problems of long operation time and low segmentation accuracy rate of the existing SAR image segmentation method and comprises the following implementation steps: 1. inputting an SAR image I to be segmented and the number c of segmentation categories; 2. carrying out Gaussian filtering on the image to obtain a filtered image X; 3. dividing the image X into a key pixel set S and a non-key pixel set L according to a local maximum pixel rule; 4. fuzzy clustering is carried out on key pixels by utilizing spatial information; 5. determining the class label of the non-key pixel by using the key pixel clustering result; 6. combining the class labels of the key pixels and the non-key pixels to obtain an intermediate segmentation result C; 7. and obtaining a final segmentation result by using the local neighborhood information smoothing result C. The SAR image segmentation method improves the accuracy of the SAR image segmentation result, reduces the segmentation time, and provides a foundation for subsequent SAR image understanding and interpretation.

Description

Rapid SAR image segmentation method based on key pixel fuzzy clustering

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to an SAR image segmentation method which can be used for image semantic recognition and image search.

Background

Synthetic aperture radar SAR images have been widely used in various fields of military and civil use in recent years due to their weather and time independent imaging processes. SAR image segmentation is a precondition and basic technology for SAR image understanding and interpretation, so that great demands are faced on accurate and quick implementation of SAR image segmentation. The SAR image segmentation is a process of dividing an SAR image into a certain number of non-overlapping uniform regions, pixels in the same region have similar characteristics, and pixels in different regions have different characteristics. However, the speckle noise widely existing in the SAR image makes the precise segmentation of the SAR image challenging, and as the size of the SAR image increases, it is important to rapidly implement the segmentation of the SAR image.

SAR image segmentation is a very critical step in SAR image processing technology, so in recent years many scholars have proposed many efficient SAR image segmentation methods, of which clustering method is a frequently used one. The main idea of the clustering method is to find a proper clustering center and then segment the image by using a certain similarity criterion. The fuzzy C-means clustering is the most widely applied clustering algorithm, the method realizes the minimization of a target function by iteratively updating a fuzzy membership matrix and a clustering center, and then each pixel in an image is divided by utilizing the fuzzy membership matrix. The traditional fuzzy clustering method only processes a single pixel point, and for the SAR image rich in speckle noise, the segmentation process is affected by very serious noise, so that the accuracy of the final segmentation result is very low.

With the improvement of the requirements of the modern society on SAR image segmentation, the traditional segmentation algorithms have the defects of low result accuracy, poor noise robustness and slow running time, so that the segmentation results obtained by the traditional method cannot meet the requirements, and researchers make some improvements on the defects, such as the improvement on the traditional fuzzy clustering: the article "Fuzzy C-means clustering with local information and kernel metric for Image segmentation" is published by Henguo et al in IEEE Transactions On Image Process,22(2014)573-584, and the article adds neighborhood items in the target function of the original Fuzzy C-means clustering, introduces local spatial distance and gray difference information, so that the robustness of the segmentation Process to noise is improved, and simultaneously adds a kernel method in the similarity measurement Process, so that the segmentation Process can further inhibit the influence of speckle noise. However, for images affected by severe speckle noise, the method is still affected by noise, resulting in lower segmentation accuracy. Meanwhile, the existing clustering method for image segmentation processes each pixel point in the image, so that the iteration process of the clustering method is very slow, and the phenomenon is more serious in the process of large-size image segmentation. Therefore, the conventional method is difficult to realize the rapid and accurate segmentation of the SAR image.

Disclosure of Invention

The invention aims to provide a rapid SAR image segmentation method based on key pixel fuzzy clustering, aiming at the defects of the prior art, so that the time for segmenting the SAR image is shortened, the influence of speckle noise on the segmentation process is inhibited, and the segmentation accuracy is improved.

In order to achieve the purpose, the technical scheme of the invention comprises the following steps:

(1) inputting an SAR image I to be segmented and the number c of segmented categories;

(2) carrying out Gaussian filtering on the SAR image I to be segmented to obtain a filtered image X;

(3) on the filtered image X, dividing all pixels in the image X into a key pixel set S and a non-key pixel set L according to a local maximum pixel rule;

(4) fuzzy clustering is carried out on key pixels in the key pixel set S to obtain a fuzzy membership matrix U and a clustering center matrix V of the key pixel set S;

(5) obtaining the class mark C of each key pixel in the key pixel set S according to the fuzzy membership matrix U_Si；

(6) According to key pixelClass label C_SiAnd clustering the center matrix V, calculating the class mark C of each non-key pixel in the non-key pixel set L_Li：

Wherein H_iRepresenting by the ith non-critical pixel L_iA central neighborhood, C_SmaxRepresentation of the content in the neighborhood H_iMiddle and non-critical pixel L_iClass label, p, of the most similar key pixel_iRepresented in the image X as non-critical pixels L_iMean value of grey values of pixels in a 5 x 5 neighborhood centered, V_kDenotes the cluster center of the kth class, n is the intersection operator,

for the null set, argmin (·) is an operator for solving the minimum value, and | · | is an operator for taking an absolute value;

(7) combined per key pixel classmark C_SiAnd a class label C for each non-critical pixel_LiObtaining an intermediate segmentation result C of the SAR image I;

(8) and smoothing the intermediate segmentation result C by using the local neighborhood information to obtain a final segmentation result of the SAR image I.

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

1. according to the SAR image clustering method based on the local maximum rule, only the key pixels selected according to the local maximum rule are clustered, so that the time required by the clustering process is shortened, the time for segmenting the SAR image is shortened, the SAR image is rapidly segmented, and compared with the method for clustering all pixels in the prior art, the SAR image clustering method based on the local maximum rule only clusters a small number of key pixels, effectively reduces the number of pixels participating in clustering, and accelerates the clustering time.

2. According to the invention, when the key pixels are clustered, the non-local information and the local information of the image are simultaneously utilized, so that the clustering process is insensitive to the influence of speckle noise, compared with an algorithm for clustering SAR images by using the local information in the prior art, the robustness to the noise is enhanced, and the accuracy of a clustering result is effectively improved.

3. The invention realizes fast and accurate segmentation by using the clustering result of the key pixels and a similarity measurement criterion of adding local information when segmenting non-key pixels, reduces the time used for segmentation, improves the robustness to noise and further improves the accuracy of the segmentation result compared with the method for clustering all pixels in the prior art.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

FIG. 2 is a graph showing the results of segmenting the FARMLAND image by using the present invention and the existing three SAR image segmentation methods of ILKFCM, NSFCM and ALFCM.

Detailed Description

The invention is described in further detail below with reference to the figures and specific embodiments.

Referring to fig. 1, the implementation steps of the invention are as follows:

step 1, inputting an SAR image I to be segmented and the number c of segmented categories.

The input SAR image I is an SAR image of an arbitrary size, the input number of divided categories c is a parameter set manually, and in the present embodiment, an SAR image of 256 × 256 in size and named FARMLAND is taken in a FARMLAND in italy, and the number of divided categories c is set to 3.

And 2, carrying out Gaussian filtering on the SAR image I to be segmented to obtain a filtered image X.

In the prior art, the filtering method for the image includes gaussian filtering, median filtering, mean filtering, and the like, and because gaussian filtering is simple to implement and has higher robustness for noise, this embodiment implements filtering on the SAR image I by using gaussian filtering to obtain the filtered image X, and the specific steps are as follows:

2a) initializing the filtered image X as an image with the size consistent with that of the SAR image I;

2b) calculated as3 x 3 Gaussian kernel template matrix Z, where the value Z of the kth element in the Gaussian kernel template matrix Z_kThe calculation is as follows:

where σ represents the variance of the Gaussian function, k₁And k₂Respectively representing the row number and the column number of the kth element in the Gaussian kernel template matrix Z, wherein exp (cndot) is a natural exponential function;

2c) calculating the gray value V of a pixel point with (i, j) in the image X according to the Gaussian kernel template matrix Z_ijComprises the following steps:

wherein P represents a pixel point I with the position coordinate in the image I as (I, j)_ijA 3 × 3 neighborhood of centers, P_kExpressing the gray value, Z, of the k-th pixel point in the neighborhood P_kRepresents the value of the kth element in the gaussian kernel template matrix Z.

2d) And generating a filtered image X according to the gray value of each pixel point obtained by calculation.

And 3, on the filtered image X, dividing all pixels in the image X into a key pixel set S and a non-key pixel set L according to a local maximum pixel rule.

The local maximum pixel rule is: taking a neighborhood R by taking each pixel point in the image as a center, if the gray value of the central pixel is the maximum in the neighborhood R, setting the central pixel as a key pixel, and specifically realizing the steps as follows:

3a) for pixel point X with position coordinate (i, j) in filtered image X_ijTaking a 3 multiplied by 3 neighborhood R with the pixel point as the center;

3b) for pixel point X according to the following formula_ijDividing:

where p ═ (i, j) denotes a pixel X_ijQ represents the position coordinate of each pixel in the neighborhood R, X (q) represents the gray value of a pixel point with the position coordinate of q in the SAR image X, and argmax (·) is an operator for taking the maximum value;

3c) and generating a key pixel set S and a non-key pixel set L according to the division result of the pixel points.

And 4, carrying out fuzzy clustering on the key pixels in the key pixel set S to obtain a fuzzy membership matrix U and a clustering center matrix V of the key pixel set S.

The existing fuzzy clustering method comprises fuzzy clustering based on local spatial information and fuzzy clustering based on non-local spatial information, and the invention adopts fuzzy clustering based on the combination of local spatial information and non-local spatial information, so that the robustness of the clustering process to noise can be improved; meanwhile, different from the clustering method in the prior art which uses all pixels for clustering, the method only clusters the key pixels in the key pixel set S, shortens the time for clustering, and comprises the following concrete steps:

4a) determining spatial locality distance key pixel S in image X_iThe nearest K key pixels are formed into a nearest neighbor set M_iIn which S is_iRepresenting the ith key pixel in the key pixel set S;

4b) computing nearest neighbor set M_iThe jth neighbor M in (2)_ijWeight w of_ij：

4b1) Calculate neighbor M as follows_ijSpatial distance weight w of_sij：

Wherein d is_ijIndicates a neighbor M_ijSpatial position of and key pixel S_iThe euclidean distance of the spatial location of (a);

4b2) calculate neighbor M as follows_ijIntensity distance weight w of_gij：

Wherein, mu_iRepresenting a key pixel S_iMean value of the grey values of the pixels in the 5X 5 neighbourhood, μ, in image X_jIndicates a neighbor M_ijIn an image X, the average value of pixel gray values in a 5 multiplied by 5 neighborhood, ln (·) is an operator of natural logarithm, exp (·) is an operator of natural exponent, and | is an operator of absolute value;

4b3) according to the neighbourhood M_ijSpatial distance weight w of_sijAnd intensity distance weight w_gijAccording to w_ij＝w_sij·w_gijComputing neighbor M_ijWeight w of_ij：

4c) Randomly initializing a fuzzy membership matrix U of a key pixel set S, and setting the initial iteration number as 1;

4d) calculating a cluster center matrix V of the current iteration, wherein the cluster center V of the kth class_kThe calculation is as follows:

where N represents the number of key pixels in the set S of key pixels, u_kiRepresenting fuzzy membership, S, of the ith key pixel to the kth class_iRepresenting the ith key pixel, and m-2 representing a fuzzy index;

4e) calculating and calculating the neighborhood item G of the ith key pixel pair of the current iteration in the kth class according to the following formula_ki：

Wherein | · | | is an operator for solving the euclidean distance;

4f) calculating a fuzzy membership matrix U of the current iteration, wherein the fuzzy membership U of the ith key pixel to the kth class_kiThe calculation is as follows:

wherein, V_jDenotes the cluster center of class j, G_jiRepresenting the neighborhood item of the ith key pixel pair of the jth class;

4g) and judging whether max | U-U1| < is true or not according to the obtained fuzzy membership matrix U of the current iteration and the fuzzy membership matrix U1 of the last iteration, if so, outputting the fuzzy membership matrix U of the current iteration and the clustering center matrix V, otherwise, adding 1 to the iteration times, and returning to the step 4d), wherein 0.00001 is a convergence threshold value.

Step 5, obtaining the class mark C of each key pixel in the key pixel set S according to the fuzzy membership matrix U_SiCalculated according to the following formula:

wherein u is_kiRepresenting the fuzzy membership of the ith key pixel to the kth class.

Step 6, according to the class mark C of the key pixel_SiAnd clustering the center matrix V, calculating the class mark C of each non-key pixel in the non-key pixel set L_Li。

This step utilizes the clustering result C of the key pixels_SiAnd the clustering center matrix V is used for determining the class mark of the non-key pixel, an iteration process is not used, the time for segmentation can be reduced, and the segmentation precision of the non-key pixel is improved. The method comprises the following concrete steps:

6a) taking non-key pixels L in the image X_iCentered neighborhood H_iWherein L is_iRepresents the ith non-critical pixel in the non-critical pixel set L;

6b) determine neighborhood H_iIf yes, executing step 6c), otherwise, executing step 6 e);

6c) calculation of Presence in H_iJ (th) key pixel H in (b)_ijAnd non-critical pixel L_iDegree of similarity a_ij：

a_ij＝a_sij·a_gij

Wherein the content of the first and second substances,

representing a key pixel H_ijAnd non-critical pixel L_iSpatial similarity of (d, d)_ijRepresenting a key pixel H_ijSpatial position of and non-critical pixels L_iThe euclidean distance of the spatial location of (a);

representing a key pixel H_ijAnd non-critical pixel L_iIntensity similarity of (a), o_iRepresenting key pixel H in image X_ijMean of pixel gray values in a 5 x 5 neighborhood centered, o_jRepresenting non-critical pixels L in an image X_iAverage of pixel gray values within a 5 x 5 neighborhood centered;

6d) obtaining the non-key pixel L according to the following formula_iClass label C_Li：

C_Li＝C_Smax

Wherein C is_SmaxTo exist in neighborhood H_iMiddle and non-critical pixel L_iDegree of similarity a_ijContinuing to execute the step 6f) for the class label of the largest key pixel;

6e) according to the clustering center matrix V, the non-key pixel L is obtained according to the following formula_iClass label C_Li:

Wherein p is_iRepresented in the image X as non-critical pixels L_iTaking the average value of the gray values of the pixels in the 5 multiplied by 5 neighborhood as the center, taking argmin (·) as an operator for solving the minimum value, and executing the step 6 f);

6f) and (4) judging whether the non-key pixels with undetermined class targets exist in the non-key pixel set L, if so, returning to the step 6a), and otherwise, executing the step 7.

Step 7, according to the obtained class mark C of each key pixel_SiAnd a class label C for each non-critical pixel_LiAnd obtaining an intermediate segmentation result C of the SAR image I.

7a) Initializing an intermediate segmentation result C into a zero matrix with the size consistent with that of the SAR image I;

7b) calculating the value C of the ith element in the intermediate segmentation result C_iComprises the following steps:

wherein, X_iAnd (4) representing the ith pixel point in the SAR image X.

And 8, smoothing the intermediate segmentation result C by using the local neighborhood information to obtain a final segmentation result of the SAR image I.

By utilizing the neighborhood information of the image, the influence of noise on the result can be reduced, the robustness of the noise is enhanced, and the accuracy of segmentation is further improved, and the method comprises the following implementation steps:

8a) according to the middle segmentation result graph C, counting the number m of the pixel points with the class mark of 1,2, a₁,m₂,...,m_c；

8b) Find m₁,m₂,...,m_cMaximum value m in_tSetting the class mark of the pixel as t;

8c) and converting the obtained class mark t into a gray value according to the t-t multiplied by 255/(c-1) to generate a final image segmentation result graph.

The effect of the invention can be further illustrated by the following simulation experiment:

1. simulation conditions are as follows:

the invention adopts Matlab R2014b software to be carried out on a computer configured as a core i52.60GHZ, a memory 8GB and a WINDOWS 7 system.

2. Simulation content:

the method and the existing segmentation methods of ILKFCM, NSFCM and ALFCM SAR images are used for segmenting the FARMLAND image, the number of segmentation categories is set to be 3, and the result is shown in FIG. 2, wherein:

fig. 2(a) is an input SAR image I to be segmented;

FIG. 2(b) is a diagram of the real segmentation result;

FIG. 2(c) is a resulting image segmented by the prior art ILKFCM method;

FIG. 2(d) is a resulting image segmented by a prior art NSFCM method;

FIG. 2(e) is a resulting image segmented with the prior ALFCM method;

FIG. 2(f) is a resulting image segmented by the method of the present invention.

As can be seen from fig. 2: compared with other methods, the segmentation result image of the method is closer to the real result image, the number of wrong points is less, the edge is clear, the noise interference is small, the segmentation accuracy is high, the robustness to noise is good, and the method provided by the invention is proved to be capable of obtaining a good segmentation result for the SAR image.

Comparing the result graphs of the present invention and the other three methods, i.e., fig. 2(c), fig. 2(d), fig. 2(E) and fig. 2(f), with the real segmentation result graph, i.e., fig. 2(b), the T and E values of the segmentation result graph of each method are counted, wherein,

t represents the number of pixels correctly divided with reference to fig. 2 (b);

e represents the number of pixels erroneously divided with reference to fig. 2 (b);

and OE-E/N represents the segmentation error rate of the used method, and SA-T/N represents the segmentation accuracy rate of the used method, wherein N represents the total number of pixels in the image.

The T and E values and the running time of the segmentation results of the method of the invention and ILKFCM, NSFCM, ALFCM methods on the FARMLAND image were counted and OE and SA values were calculated as shown in table 1. Wherein lower OE values indicate better results, higher SA values indicate better results, and lower run times indicate better results.

TABLE 1 results of the segmentation of FARMLAND by different methods

Method of producing a composite material	SA	OE	time
				ILKFCM	0.7049	0.2951	298.3544s
NSFCM	0.6851	0.3149	168.5366s
				ALFCM	0.5648	0.4352	112.0980s
The invention	0.7973	0.2027	49.6627s

As can be seen from table 1, for the SAR image FARMLAND, the segmentation result of the ALFCM method is the worst, the method has the lowest segmentation accuracy SA and the highest segmentation error rate OE, the segmentation accuracy SA of the ILKFCM and the NSFCM method is not high, and the running time of the ILKFCM segmentation is the longest and is far longer than the time used in the present invention. The method has the lowest segmentation error rate OE and the highest accuracy rate SA which are higher than SA values of other three methods by more than 9 percent, and the running time of the method is at least 60 seconds faster than that of the other three methods, which shows that the SAR image can be rapidly segmented by adopting the method to obtain more accurate results.

Claims

1. A fast SAR image segmentation method based on key pixel fuzzy clustering comprises the following steps:

(6) Class label C based on key pixel_SiAnd clustering the center matrix V, calculating the class mark C of each non-key pixel in the non-key pixel set L_Li：

Wherein H_iRepresenting by the ith non-critical pixel L_iA central neighborhood, C_SmaxRepresentation of the content in the neighborhood H_iMiddle and non-critical pixel L_iClass label, p, of the most similar key pixel_iRepresented in the image X as non-critical pixels L_iMean value of grey values of pixels in a 5 x 5 neighborhood centered, V_gRepresenting the cluster center of the g-th class, n is the intersection operator,

for the null set, argmin (·) is an index operator for solving the minimum value, and | · | is an operator for taking an absolute value;

2. The method according to claim 1, characterized in that step (2) is performed by gaussian filtering the SAR image I to be segmented, as follows:

2b) calculating a Gaussian kernel template matrix Z of size 3 x 3, wherein the value Z of the b-th element in the Gaussian kernel template matrix Z_bThe calculation is as follows:

where σ represents the variance of the Gaussian function, b₁And b₂Respectively representing the row number and the column number of the b-th element in the Gaussian core template matrix Z, wherein exp (cndot) is a natural exponential function;

2c) according to the Gaussian kernel template matrix Z, calculating the gray value D of the pixel point with the position (i, j) in the image X_ijComprises the following steps:

wherein P represents a pixel point I with the position coordinate in the image I as (I, j)_ijA 3 × 3 neighborhood of centers, P_bExpressing the gray value of the b-th pixel point in the neighborhood P;

3. The method according to claim 1, wherein the step (3) of dividing all pixels in the image X into a key pixel set S and a non-key pixel set L according to a local maximum pixel rule on the filtered image X is performed as follows:

3b) for pixel point X according to the following formula_ijDividing:

where p ═ (i, j) denotes a pixel X_ijQ represents the position coordinate of each pixel in the neighborhood R, X (q) represents the gray value of a pixel point with the position coordinate of q in the SAR image X, and argmax (·) is an index value operator taking the maximum value;

4. The method according to claim 1, wherein the step (4) performs fuzzy clustering on the key pixels in the key pixel set S according to the following steps:

4b1) Calculate neighbor M as follows_ijSpatial distance weight w of_sij：

4b2) calculate neighbor M as follows_ijIntensity distance weight w of_gij：

Wherein, mu_iRepresenting a key pixel S_iMean value of the grey values of the pixels in the 5X 5 neighbourhood, μ, in image X_jIndicates a neighbor M_ijThe average value of the gray values of pixels in 5 multiplied by 5 neighborhoods in the image X, wherein ln (·) is an operator of taking a natural logarithm;

4b3) according to the neighbourhood M_ijSpatial distance weight w of_sijAnd intensity distance weight w_gijAccording to w_ij＝w_sij·w_gijComputing neighbor M_ijWeight w of_ij；

4d) calculating a cluster center matrix V of the current iteration, wherein the cluster center V of the g-th class_gThe calculation is as follows:

where N represents the number of key pixels in the set S of key pixels, u_giRepresenting the fuzzy membership degree of the ith key pixel to the g-th class, wherein t is 2 to represent a fuzzy index;

4e) calculating the neighborhood item G of the ith key pixel pair of the current iteration in the category G according to the following formula_gi：

Wherein | · | | is an operator for solving the euclidean distance;

4f) calculating a fuzzy membership matrix U of the current iteration, wherein the fuzzy membership U of the ith key pixel to the g class_giThe calculation is as follows:

wherein, V_gDenotes the cluster center of the G-th class, G_giRepresenting a neighborhood item of the ith key pixel pair of the g type;

5. The method according to claim 1, wherein the step (5) obtains the class label C of each key pixel in the key pixel set S according to the fuzzy membership matrix U_SiCalculated according to the following formula:

wherein u is_giAnd (3) representing the fuzzy membership degree of the ith key pixel to the g class, wherein argmax (·) is an operator for solving the maximum value.

6. The method of claim 1, wherein the step (7) combines the class labels C for each key pixel_SiAnd a class label C for each non-critical pixel_LiAnd obtaining an intermediate segmentation result C of the SAR image I, and performing the following steps:

wherein, X_iAnd (4) representing the ith pixel point in the SAR image X.

7. The method according to claim 1, characterized in that step (8) utilizes local neighborhood information to smooth the intermediate segmentation result C to obtain a final segmentation result of the SAR image I, and the method comprises the following steps:

8a) according to the intermediate segmentation result C, counting the number m of the pixel points with the class mark of 1,2, the₁,m₂,...,m_c；

8b) Find m₁,m₂,...,m_cMaximum value m in_gSetting the class label of the pixel as g;

8c) and converting the obtained class mark g into a gray value according to the g-g multiplied by 255/(c-1) to generate a final image segmentation result graph.