CN103020978B - SAR (synthetic aperture radar) image change detection method combining multi-threshold segmentation with fuzzy clustering - Google Patents

Abstract

The invention discloses an SAR (synthetic aperture radar) image change detection method combining multi-threshold segmentation with fuzzy clustering. The method mainly aims at overcoming the defect of existing fuzzy clustering algorithms and is used for SAR image change detection by combining multi-threshold segmentation with fuzzy clustering. The implementation steps of the method include: (1), subjecting two SAR images to median filtering; (2), calculating to obtain a logarithmic ratio differential image after normalization; (3), adopting the Otsu method based on standard particle swarm optimization to perform multi-threshold segmentation to the logarithmic ratio differential image after normalization; (4), initializing membership matrixes U0 and U1; (5), adopting the FLICM (fuzzy local information C-means) algorithm to perform fuzzy clustering to pixels which cannot be determined whether changes occur or not after multi-threshold segmentation; (6), deblurring; and (7), outputting change detection results. The multi-threshold segmentation and fuzzy clustering are combined for SAR image change detection, so that change detection time is reduced, and change detection accuracy is improved.

Description

In conjunction with the SAR image change detection of multi-threshold segmentation and fuzzy clustering

Technical field

The invention belongs to field of computer technology, further relate to the SAR image change detection in conjunction with multi-threshold segmentation and fuzzy clustering in technical field of image processing.The present invention obtains differential image by the SAR image of the different phase of two width, again multi-threshold segmentation and fuzzy clustering are carried out to differential image, realize SAR image change and detect, can be used for atural object and cover and utilization, Natural calamity monitoring and assessment, city planning, the fields such as map rejuvenation.

Background technology

Synthetic-aperture radar (Synthetic Aperture Radar, SAR) technology is more and more widely used in recent years.SAR image is compared with ordinary optical remote sensing images, and having can the feature of round-the-clock acquisition, and improves constantly along with SAR image resolution, and the image processing techniques based on it also gets more and more.The change detection techniques of image refers to by the contrast and analysis of areal at two width remote sensing images of different time, obtains the landforms change information of this area in these two periods.Change detection techniques at present based on SAR image has obtained many application, such as in Forest cover change, Natural calamity monitoring and assessment etc.

For the region how finding out change from two width SAR image, roughly having two kinds of methods: one is classification and predicting method, is exactly first classify separately to different SAR image, and then the result after match stop, finds out the region of change; Another method is then the differential image first obtaining them from different SAR image, then analyzes differential image, thus finds region of variation, and a kind of rear method is applied more at present.Will carry out a kind of division of two classifications exactly to it to the process reality of differential image, therefore clustering algorithm can by the classification problem solved differential image.

Fuzzy C-mean algorithm (Fuzzy C-Means, FCM) algorithm is one of most popular clustering algorithm, it adopts fuzzy theory to set up the uncertainty description of sample class, obtains the degree of membership of sample to all cluster centres, and then determine the classification of each sample by optimization object function.But because FCM does not consider the spatial information of sample, so more responsive to noise ratio, have impact on the effect of cluster.Fuzzy local message C average (Fuzzy Local InformationC-Means, FLICM) algorithm is a kind of newer clustering algorithm, it has done improvement for FCM to the more sensitive shortcoming of noise ratio, the local spatial information of sample is introduced in the objective function of cluster, achieve Clustering Effect more better than FCM, but owing to both increasing the calculating of neighborhood information to all samples, compared with FCM, its operand is larger.

Patent " SAR image change detection based on the quantum-inspired immune clone " (number of patent application 201010230980.6 of Xian Electronics Science and Technology University's application, publication number CN101908213A), the method is by quantum bit definitions cluster centre, the cluster centre that search is optimum also obtains global threshold, but the deficiency that the method exists is, the objective function only employing FCM, to construct antibody parent fitness function, does not consider the neighborhood information of pixel.

Summary of the invention

The object of the invention is to the deficiency overcoming the existence of above-mentioned prior art, multi-threshold segmentation and fuzzy clustering are combined, detect for SAR image change, change detection time to reduce, and improve the precision changing and detect.

For achieving the above object, the present invention, when carrying out change and detecting, first utilizes multi-threshold segmentation to determine the classification of " simply " pixel, then all the other " difficulty " pixels is left for FLICM fuzzy clustering algorithm to differentiate.

The present invention includes following steps:

(1) medium filtering

Choose 3 × 3 conventional median filters and pre-service is carried out to two SAR image to be detected, obtain two width images after medium filtering;

(2) the logarithm ratio differential image after normalization is obtained

2a) adopt logarithm ratio difference formula, by two width images after medium filtering, obtain logarithm ratio differential image;

The logarithm ratio differential image after normalization is obtained after 2b) adopting normalization formula to be normalized logarithm ratio differential image;

(3) multi-threshold segmentation

Adopt the Da-Jin algorithm optimized based on standard particle group, carry out multi-threshold segmentation to the logarithm ratio differential image after normalization, the pixel after segmentation in image is divided into three classes: the pixel do not changed, there occurs change pixel and can not determine the pixel that whether there occurs change;

(4) initialization subordinated-degree matrix

4a) initialization subordinated-degree matrix U ₀, U ₀the degree of membership being under the jurisdiction of non-changing class of storage pixel, a degree of membership real number in interval [0,1] upper value represents, belongs to the pixel that do not change at U after multi-threshold segmentation ₀in degree of membership get 1, belong to after multi-threshold segmentation there occurs change pixel at U ₀in degree of membership get 0, belong to after multi-threshold segmentation and can not determine that the pixel that whether there occurs change is at U ₀in degree of membership then stochastic generation;

4b) initialization subordinated-degree matrix U ₁, U ₁the degree of membership being under the jurisdiction of change class of storage pixel, the degree of membership real number that is gone up value in interval [0,1] represents, belongs to the pixel that do not change at U after multi-threshold segmentation ₁in degree of membership get 0, belong to after multi-threshold segmentation there occurs change pixel at U ₁in degree of membership get 1, belong to after multi-threshold segmentation and can not determine that the pixel that whether there occurs change is at U ₁in degree of membership then equal 1 and deduct this pixel at U ₀in degree of membership;

(5) fuzzy clustering

Adopt FLICM algorithm to can not determine that the pixel that whether there occurs change carries out fuzzy clustering to belonging to after multi-threshold segmentation, iteration upgrades until reach predetermined end condition, and namely before and after iteration, the maximum knots modification of degree of membership is less than 0.00001;

(6) deblurring

According to each pixel in the logarithm ratio differential image after normalization be under the jurisdiction of non-changing class and change class be subordinate to angle value, pixel is judged to the ownership of and is subordinate to that larger class of angle value, thus all pixels in the logarithm ratio differential image after normalization be divide into two classes: non-changing class and change class;

(7) exporting change testing result.

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

1. first the present invention utilizes multi-threshold segmentation to determine the classification of " simply " pixel, again all the other " difficulty " pixels are left for FLICM algorithm to differentiate, directly carry out cluster to pixel with existing FLICM algorithm to compare with differentiation, present invention reduces operand and the working time of change detection.

2. the present invention is when using FLICM algorithm to carry out fuzzy clustering, local window N _iin the classification of some non-central pixel determine in the multi-threshold segmentation stage, these neighborhood informations determined improve the effect that prior art carries out fuzzy clustering, the precision making the present invention further increase change to detect.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 carries out changing the result detected to Ottawa area floods SAR image in emulation experiment;

Fig. 3 carries out changing the result detected to Mexico's area fire SAR image in emulation experiment;

Fig. 4 carries out changing the result detected to Bern Urban flood SAR image in emulation experiment.

Embodiment

With reference to Fig. 1, comprise the steps: in conjunction with the SAR image change detection of multi-threshold segmentation and fuzzy clustering in the present invention

(1) medium filtering

Choose 3 × 3 conventional median filters and pre-service is carried out to two SAR image to be detected, obtain two width images after medium filtering;

(2) the logarithm ratio differential image after normalization is obtained

2a) adopt following logarithm ratio difference formula, by two width images after medium filtering, obtain logarithm ratio differential image:

I ₃＝|log(I ₁+1)-log(I ₂+1)|

Wherein, I ₃represent the grey scale pixel value of logarithm ratio differential image, I ₁and I ₂represent the grey scale pixel value of two width images after medium filtering respectively;

The logarithm ratio differential image after normalization is obtained after 2b) adopting following normalization formula to be normalized logarithm ratio differential image:

$I_D=\frac{255\×(I_3-I_{\min })}{(I_{\max }-I_{\min })}$

Wherein, I _drepresent the grey scale pixel value of the logarithm ratio differential image after normalization, I ₃represent the grey scale pixel value of logarithm ratio differential image, I _maxrepresent I ₃in maximum gray-scale value, I _minrepresent I ₃in minimum gray-scale value;

(3) multi-threshold segmentation

Adopt the Da-Jin algorithm optimized based on standard particle group to carry out multi-threshold segmentation to the logarithm ratio differential image after normalization, the pixel after segmentation in image is divided into three classes: the pixel do not changed, there occurs change pixel and can not determine the pixel that whether there occurs change;

The concrete steps of multi-threshold segmentation are as follows:

3a) the n dimension particle that n the threshold coding being used for splitting image is during standard particle group optimizes;

3b) use n the threshold value that standard particle group Optimizing Search makes Da-Jin algorithm objective function maximum, Da-Jin algorithm objective function is as follows:

Maximize σ ²＝ω ₀(μ ₀-μ _T) ²+ω ₁(μ ₁-μ _T) ²+...+ω _n(μ _n-μ _T) ²

Logarithm ratio differential image after normalization can be divided into n+1 part by n threshold value, in above formula, and σ ²represent the variance between this n+1 part, ω ₀, ω ₁..., ω _nrepresent the probability that various piece occurs respectively, μ ₀, μ ₁..., μ _nrepresent the average gray of various piece respectively, μ _trepresent the average gray of whole image;

In standard particle group optimizes, the more new formula of particle d dimension flying speed is:

$v_i^d=\mathrm{\ω}{v}_i^d+c_1{\mathrm{rand}}_1^d({\mathrm{pBest}}_i^d-x_i^d)+c_2{\mathrm{rand}}_2^d({\mathrm{gBest}}^d-x_i^d)$

Wherein, represent that the d of i-th particle in population ties up flying speed, ω represents inertia weight, c ₁and c ₂represent Studying factors, represent that the d of i-th particle in population ties up position, with two separate random numbers be uniformly distributed between 0 and 1, pBest _ibe the history optimal location of i-th particle, gBest is the history optimal location of whole population;

In standard particle group optimizes, the more new formula of particle d dimension position is:

$x_i^d=x_i^d+v_i^d$

Wherein, represent that the d of i-th particle in population ties up position, represent that the d of i-th particle in population ties up flying speed;

3c) by the search that standard particle group optimizes, find n the threshold value making Da-Jin algorithm objective function maximum, this n threshold value is divided into n+1 part the logarithm ratio differential image after normalization, is designated as: C according to gray-scale value order from small to large ₀, C ₁..., C _n.C ₀in pixel be classified as the pixel do not changed, C _nin pixel be classified as the pixel that there occurs change, and C ₁to C _n-1in pixel be classified as the pixel can not determine and whether there occurs change, thus the pixel in the logarithm ratio differential image after normalization be divide into three classes;

(4) initialization subordinated-degree matrix

4a) initialization subordinated-degree matrix U ₀, U ₀the degree of membership being under the jurisdiction of non-changing class of storage pixel, a degree of membership real number in interval [0,1] upper value represents, belongs to the pixel that do not change at U after multi-threshold segmentation ₀in degree of membership get 1, belong to after multi-threshold segmentation there occurs change pixel at U ₀in degree of membership get 0, belong to after multi-threshold segmentation and can not determine that the pixel that whether there occurs change is at U ₀in degree of membership then stochastic generation;

4b) initialization subordinated-degree matrix U ₁, U ₁the degree of membership being under the jurisdiction of change class of storage pixel, the degree of membership real number that is gone up value in interval [0,1] represents, belongs to the pixel that do not change at U after multi-threshold segmentation ₁in degree of membership get 0, belong to after multi-threshold segmentation there occurs change pixel at U ₁in degree of membership get 1, belong to after multi-threshold segmentation and can not determine that the pixel that whether there occurs change is at U ₁in degree of membership then equal 1 and deduct this pixel at U ₀in degree of membership;

(5) fuzzy clustering

Adopt FLICM algorithm to can not determine that the pixel that whether there occurs change carries out fuzzy clustering to belonging to after multi-threshold segmentation, iteration upgrades until reach predetermined end condition, and namely before and after iteration, the maximum knots modification of degree of membership is less than 0.00001;

5a) FLICM algorithm uses degree of membership to represent that pixel may belong to the degree of certain classification, and it considers gradation of image information and space neighborhood information in cluster process simultaneously, and the objective function of FLICM is as follows:

$J=\underset{i\∈N}{\mathrm{\Σ}}\underset{k=1}{\overset{C}{\mathrm{\Σ}}}[u_{\mathrm{ki}}^m{d}^2(x_i,v_k)+G_{\mathrm{ki}}]$

Wherein, J represents target function value, and N belongs to the set that can not determine the pixel composition that whether there occurs change after representing multi-threshold segmentation, C to be value be 2 class number, u _kirepresent the degree of membership of i-th pixel in kth class, x _irepresent the gray-scale value of i-th pixel, v _krepresent the cluster centre of kth class, d ²(x _i, v _k) representing the distance metric of i-th pixel to the cluster centre of kth class, m represents that value is the FUZZY WEIGHTED index of 2, G _kirepresent the fuzzy coefficient between i-th pixel and the cluster centre of kth class, shown in it is defined as follows:

$G_{\mathrm{ki}}=\underset{\underset{i\≠j}{j\∈N_i}}{\mathrm{\Σ}}\frac{1}{e_{\mathrm{ij}}+1}{(1-u_{\mathrm{kj}})}^m{d}^2(x_j,v_k)$

Wherein, G _kirepresent the fuzzy coefficient between i-th pixel and the cluster centre of kth class, N _irepresent that the size centered by i-th pixel is the local window of 3 × 3, x _jrepresent the gray-scale value of a jth pixel and this pixel falls into local window N _iin non-central pixel, v _krepresent the cluster centre of kth class, u _kjrepresent the degree of membership of a jth pixel in kth class, e _ijrepresent the space Euclidean distance between i-th pixel and a jth pixel, d ²(x _j, v _k) representing the distance metric of a jth pixel to the cluster centre of kth class, m represents that value is the FUZZY WEIGHTED index of 2;

5b) mode of FLICM algorithm employing iterative computation is obtained and is made its objective function minimum a group be subordinate to angle value, and in FLICM algorithm, the iterative computation formula of degree of membership is as follows:

$u_{\mathrm{ki}}=\frac{1}{\underset{t=1}{\overset{C}{\mathrm{\Σ}}}{\left(\frac{d^2(x_i,v_k)+G_{\mathrm{ki}}}{d^2(x_i,v_t)+G_{\mathrm{ti}}}\right)}^{\frac{1}{m-1}}}$

Wherein, u _kirepresent the degree of membership of i-th pixel in kth class, C to be value be 2 class number, x _irepresent the gray-scale value of i-th pixel, v _krepresent the cluster centre of kth class, v _trepresent the cluster centre of t class, G _kirepresent the fuzzy coefficient between i-th pixel and the cluster centre of kth class, G _tirepresent i-th fuzzy coefficient between pixel and the cluster centre of t class, d ²(x _i, v _k) represent the distance metric of i-th pixel to the cluster centre of kth class, d ²(x _i, v _t) representing the distance metric of i-th pixel to the cluster centre of t class, m represents that value is the FUZZY WEIGHTED index of 2;

5c) in FLICM algorithm, the iterative computation formula of cluster centre is as follows:

$v_k=\frac{\underset{i\∈N}{\mathrm{\Σ}}{u}_{\mathrm{ki}}^m{x}_i}{\underset{i\∈N}{\mathrm{\Σ}}{u}_{\mathrm{ki}}^m}$

Wherein, v _krepresent the cluster centre of kth class, N belongs to the set that can not determine the pixel composition that whether there occurs change after representing multi-threshold segmentation, u _kirepresent the degree of membership of i-th pixel in kth class, x _irepresent the gray-scale value of i-th pixel, m represents that value is the FUZZY WEIGHTED index of 2;

(6) deblurring

According to each pixel in the logarithm ratio differential image after normalization be under the jurisdiction of non-changing class and change class be subordinate to angle value, pixel is judged to the ownership of and is subordinate to that larger class of angle value, thus all pixels in the logarithm ratio differential image after normalization be divide into two classes: non-changing class and change class;

(7) exporting change testing result.

Effect of the present invention can be further illustrated by following emulation experiment:

1. emulation experiment condition

Emulation experiment environment: operating system is Windows XP, CPU is AMD Athlon1.60GHz, inside saves as 1.75GB, programming platform is Visual C++6.0.

Emulation experiment one is that the SAR image change of Ottawa area floods detects, image size is 290 × 350, Fig. 2 (a) is the SAR image in May, 1997, Fig. 2 (b) is the SAR image in August, 1997, Fig. 2 (c) is that corresponding change detects reference diagram, Fig. 2 (d) is the change testing result of FCM algorithm, and Fig. 2 (e) is the change testing result of FLICM algorithm, and Fig. 2 (f) is change testing result of the present invention.

Emulation experiment two is that the SAR image change of Mexico's area fire detects, image size is 512 × 512, Fig. 3 (a) is the SAR image in April, 2000, Fig. 3 (b) is the SAR image in May, 2002, Fig. 3 (c) is that change detects reference diagram, Fig. 3 (d) is the change testing result of FCM algorithm, and Fig. 3 (e) is the change testing result of FLICM algorithm, and Fig. 3 (f) is change testing result of the present invention.

Emulation experiment three is that the SAR image change of Bern Urban flood detects, image size is 301 × 301, Fig. 4 (a) is the SAR image in April, 1999, Fig. 4 (b) is the SAR image in May, 1999, Fig. 4 (c) is that change detects reference diagram, Fig. 4 (d) is the change testing result of FCM algorithm, and Fig. 4 (e) is the change testing result of FLICM algorithm, and Fig. 4 (f) is change testing result of the present invention.

2. emulation experiment content and result

Optimum configurations in emulation experiment is as follows: in the present invention, the threshold number n of multi-threshold segmentation is set to 3 (providing according in the threshold number Parameter analysis emulation experiment below of selection); Two Studying factors of Plays particle group optimizing of the present invention are all set to 2.0, and inertia weight is set to along with search procedure linearly reduces from 0.9 to 0.4, and standard particle group optimize end condition be objective function evaluate number of times reach 5000 times; The end condition of the fuzzy clustering process in emulation experiment in three kinds of methods is all set to the maximum knots modification of degree of membership before and after iteration and is less than 0.00001.Data in following emulation experiment are the average result after 20 independent operatings.

Emulation experiment one: three kind of method to the change testing result of this problem as shown in Fig. 2 (d) to (f), to the concrete comparative analysis of testing result in table 1.As can be seen from Figure 2, testing result of the present invention and reference diagram are the most close.As can be seen from Table 1, the undetected pixel count of the present invention is fewer than FCM and FLICM respectively 1246 and 1309, and total erroneous pixel number is also few than both respectively 1003 and 372.Simultaneously upper as can be seen from working time, although invention increases the process of multi-threshold segmentation, after thereby also reducing, carry out the task amount of fuzzy clustering, so working time is still faster much than directly using FLICM algorithm to carry out cluster.

Table 1Ottawa area floods change testing result

Emulation experiment two: three kinds of methods to the change testing result of this problem as shown in Fig. 3 (d) to (f), to the concrete comparative analysis of testing result in table 2.As can be seen from Table 2, the undetected pixel count of the present invention is fewer than FCM and FLICM respectively 623 and 530, and total erroneous pixel number is also few than both respectively 149 and 383.The present invention, by only considering that the multi-threshold segmentation of half-tone information and emphasis consider that the FLICM algorithm of neighborhood information organically combines, improves noiseproof feature, greatly reduces undetected number, makes overall detection perform be better than two kinds of methods contrasted.Be about 30% of FLICM algorithm working time of the present invention as can be seen from Table 2 simultaneously.

Table 2 Mexico area fire change testing result

Emulation experiment three: three kinds of methods to the change testing result of this problem as shown in Fig. 4 (d) to (f), to the concrete comparative analysis of testing result in table 3.As can be seen from Table 3, for this change test problems, the present invention undetected several in still there is obvious advantage, thus total erroneous pixel number is also less, and only have 112, false detection rate only has 0.12%, and working time is also very short, less than 25% of FLICM algorithm.

Table 3Bern Urban flood change testing result

Use following formula can calculate the present invention and on average reduce about 70% than FLICM algorithm the working time in above-mentioned three groups of emulation experiments:

$T=\frac{T_2-T_1}{T_2}$

Wherein, the minimizing ratio of T represents the present invention compared with FLICM algorithm working time, T ₁represent working time of the present invention, T ₂represent the working time of FLICM algorithm.

3. the Parameter analysis emulation experiment of threshold number

The selection carrying out threshold number n during multi-threshold segmentation in the present invention has considerable influence to the effect that change detects.In order to reasonably select the size of threshold number n, carried out Parameter analysis to threshold number n, independent operating 10 the inventive method under different threshold number, the testing result after average is as shown in table 4.As can be seen from Table 4, the too small meeting of threshold number n obviously increases the false detection rate that change detects; Threshold number n is excessive, can cause the increase of working time.After considering, it is more reasonably select that threshold number n gets 3.

The Parameter analysis the simulation experiment result of the threshold number n in table 4 the present invention

Claims (4)

1., in conjunction with a SAR image change detection for multi-threshold segmentation and fuzzy clustering, comprise the steps:

(1) medium filtering

Choose 3 × 3 median filters and pre-service is carried out to two SAR image to be detected, obtain two width images after medium filtering;

(2) the logarithm ratio differential image after normalization is obtained

2a) adopt logarithm ratio difference formula, by two width images after medium filtering, obtain logarithm ratio differential image;

The logarithm ratio differential image after normalization is obtained after 2b) adopting normalization formula to be normalized logarithm ratio differential image;

(3) multi-threshold segmentation

Adopt the Da-Jin algorithm optimized based on standard particle group, carry out multi-threshold segmentation to the logarithm ratio differential image after normalization, the pixel after segmentation in image is divided into three classes: the pixel do not changed, there occurs change pixel and can not determine the pixel that whether there occurs change;

The concrete steps that the described Da-Jin algorithm optimized based on standard particle group carries out multi-threshold segmentation to the logarithm ratio differential image after normalization are as follows:

The first step is a n dimension particle during standard particle group optimizes n the threshold coding being used for splitting image;

Second step, use n the threshold value that standard particle group Optimizing Search makes Da-Jin algorithm objective function maximum, Da-Jin algorithm objective function is as follows:

Maximize σ ²＝ω ₀(μ ₀-μ _T) ²+ω ₁(μ ₁-μ _T) ²+...+ω _n(μ _n-μ _T) ²

Logarithm ratio differential image after normalization can be divided into n+1 part by n threshold value, in above formula, and σ ²represent the variance between this n+1 part, ω ₀, ω ₁..., ω _nrepresent the probability that various piece occurs respectively, μ ₀, μ ₁..., μ _nrepresent the average gray of various piece respectively, μ _trepresent the average gray of whole image;

In standard particle group optimizes, the more new formula of particle d dimension flying speed is:

$v_i^d=\mathrm{\ω}{v}_i^d+c_1{\mathrm{rand}}_1^d(p{\mathrm{Best}}_i^d-x_i^d)+c_2{\mathrm{rand}}_2^d(g{\mathrm{Best}}^d-x_i^d)$

Wherein, represent that the d of i-th particle in population ties up flying speed, ω represents inertia weight, c ₁and c ₂represent Studying factors, represent that the d of i-th particle in population ties up position, with two separate random numbers be uniformly distributed between 0 and 1, pBest _ibe the history optimal location of i-th particle, gBest is the history optimal location of whole population;

In standard particle group optimizes, the more new formula of particle d dimension position is:

$x_i^d=x_i^d+v_i^d$

Wherein, represent that the d of i-th particle in population ties up position, represent that the d of i-th particle in population ties up flying speed;

3rd step, n threshold value is divided into n+1 part the logarithm ratio differential image after normalization, is designated as: C according to gray-scale value order from small to large ₀, C ₁..., C _n; C ₀in pixel be classified as the pixel do not changed, C _nin pixel be classified as the pixel that there occurs change, and C ₁to C _n-1in pixel be classified as the pixel can not determine and whether there occurs change, thus the pixel in the logarithm ratio differential image after normalization be divide into three classes;

(4) initialization subordinated-degree matrix

4a) initialization subordinated-degree matrix U ₀, U ₀storage pixel is under the jurisdiction of the degree of membership of non-changing class, and a degree of membership real number in interval [0,1] upper value represents, belongs to the pixel that do not change at U after multi-threshold segmentation ₀in degree of membership get 1, belong to after multi-threshold segmentation there occurs change pixel at U ₀in degree of membership get 0, belong to after multi-threshold segmentation and can not determine that the pixel that whether there occurs change is at U ₀in degree of membership then stochastic generation;

4b) initialization subordinated-degree matrix U ₁, U ₁storage pixel is under the jurisdiction of the degree of membership of change class, and the degree of membership real number that is gone up value in interval [0,1] represents, belongs to the pixel that do not change at U after multi-threshold segmentation ₁in degree of membership get 0, belong to after multi-threshold segmentation there occurs change pixel at U ₁in degree of membership get 1, belong to after multi-threshold segmentation and can not determine that the pixel that whether there occurs change is at U ₁in degree of membership then equal 1 and deduct this pixel at U ₀in degree of membership;

(5) fuzzy clustering

Adopt FLICM algorithm to can not determine that the pixel that whether there occurs change carries out fuzzy clustering to belonging to after multi-threshold segmentation, iteration upgrades until reach predetermined end condition, and namely before and after iteration, the maximum knots modification of degree of membership is less than 0.00001;

(6) deblurring

Be under the jurisdiction of the degree of membership of non-changing class according to each pixel in the logarithm ratio differential image after normalization and be under the jurisdiction of the degree of membership of change class, pixel is judged to the ownership of that class that degree of membership is larger, thus all pixels in the logarithm ratio differential image after normalization be divide into two classes: non-changing class and change class;

(7) exporting change testing result.

2. the SAR image change detection in conjunction with multi-threshold segmentation and fuzzy clustering according to claim 1, is characterized in that, step 2a) described in logarithm ratio difference formula as follows:

I ₃＝|log(I ₁+1)-log(I ₂+1)|

Wherein, I ₃represent the grey scale pixel value of logarithm ratio differential image, I ₁and I ₂represent the grey scale pixel value of two width images after medium filtering respectively.

3. the SAR image change detection in conjunction with multi-threshold segmentation and fuzzy clustering according to claim 1, is characterized in that, step 2b) described in normalization formula as follows:

$I_D=\frac{255\×(I_3-I_{\min })}{(I_{\max }-I_{\min })}$

Wherein, I _drepresent the grey scale pixel value of the logarithm ratio differential image after normalization, I ₃represent the grey scale pixel value of logarithm ratio differential image, I _maxrepresent I ₃in maximum gray-scale value, I _minrepresent I ₃in minimum gray-scale value.

4. the SAR image change detection in conjunction with multi-threshold segmentation and fuzzy clustering according to claim 1, is characterized in that, being implemented as follows of the FLICM fuzzy clustering algorithm described in step (5):

FLICM algorithm uses degree of membership to represent that pixel may belong to the degree of certain classification, and it considers gradation of image information and space neighborhood information in cluster process simultaneously, and the objective function of FLICM is as follows:

$J=\underset{i\∈N}{\mathrm{\Σ}}\underset{k=1}{\overset{C}{\mathrm{\Σ}}}[u_{\mathrm{ki}}^m{d}^2(x_i,v_k)+G_{\mathrm{ki}}]$

Wherein, J represents target function value, and N belongs to the set that can not determine the pixel composition that whether there occurs change after representing multi-threshold segmentation, C to be value be 2 class number, u _kirepresent the degree of membership of i-th pixel in kth class, x _irepresent the gray-scale value of i-th pixel, v _krepresent the cluster centre of kth class, d ²(x _i, v _k) representing the distance metric of i-th pixel to the cluster centre of kth class, m represents that value is the FUZZY WEIGHTED index of 2, G _kirepresent the fuzzy coefficient between i-th pixel and the cluster centre of kth class, shown in it is defined as follows:

$G_{\mathrm{ki}}=\underset{i\≠j}{\underset{j\∈N_i}{\mathrm{\Σ}}}\frac{1}{e_{\mathrm{ij}}+1}{(1-u_{\mathrm{kj}})}^m{d}^2(x_j,v_k)$

Wherein, G _kirepresent the fuzzy coefficient between i-th pixel and the cluster centre of kth class, N _irepresent that the size centered by i-th pixel is the local window of 3 × 3, x _jrepresent the gray-scale value of a jth pixel and this pixel falls into local window N _iin non-central pixel, v _krepresent the cluster centre of kth class, u _kjrepresent the degree of membership of a jth pixel in kth class, e _ijrepresent the space Euclidean distance between i-th pixel and a jth pixel, d ²(x _j, v _k) representing the distance metric of a jth pixel to the cluster centre of kth class, m represents that value is the FUZZY WEIGHTED index of 2;

The mode of FLICM algorithm employing iterative computation is obtained and is made its objective function minimum a group be subordinate to angle value, and in FLICM algorithm, the iterative computation formula of degree of membership is as follows:

$u_{\mathrm{ki}}=\frac{1}{\underset{t=1}{\overset{C}{\mathrm{\Σ}}}{\left(\frac{d^2(x_i,v_k)+G_{\mathrm{ki}}}{d^2(x_i,v_t)+G_{\mathrm{ti}}}\right)}^{\frac{1}{m-1}}}$

Wherein, u _kirepresent the degree of membership of i-th pixel in kth class, C to be value be 2 class number, x _irepresent the gray-scale value of i-th pixel, v _krepresent the cluster centre of kth class, v _trepresent the cluster centre of t class, G _kirepresent the fuzzy coefficient between i-th pixel and the cluster centre of kth class, G _tirepresent i-th fuzzy coefficient between pixel and the cluster centre of t class, d ²(x _i, v _k) represent the distance metric of i-th pixel to the cluster centre of kth class, d ²(x _i, v _t) representing the distance metric of i-th pixel to the cluster centre of t class, m represents that value is the FUZZY WEIGHTED index of 2;

In FLICM algorithm, the iterative computation formula of cluster centre is as follows:

$v_k=\frac{\underset{i\∈N}{\mathrm{\Σ}}{u}_{\mathrm{ki}}^m{x}_i}{\underset{i\∈N}{\mathrm{\Σ}}{u}_{\mathrm{ki}}^m}$

Wherein, v _krepresent the cluster centre of kth class, N belongs to the set that can not determine the pixel composition that whether there occurs change after representing multi-threshold segmentation, u _kirepresent the degree of membership of i-th pixel in kth class, x _irepresent the gray-scale value of i-th pixel, m represents that value is the FUZZY WEIGHTED index of 2;

Use FLICM algorithm to can not determine that the pixel that whether there occurs change carries out fuzzy clustering to belonging to after multi-threshold segmentation, until reach predetermined end condition, namely before and after iteration, the maximum knots modification of degree of membership is less than 0.00001.