CN103020977B - SAR (synthetic aperture radar) segmentation method based on polychotomy weighting segmentation - Google Patents

Abstract

The invention discloses an SAR (synthetic aperture radar) segmentation method based on polychotomy weighting segmentation and mainly overcomes the defect that image segmentation information extracted is incomprehensive and low in segmentation quality. The implementation process for the method includes firstly, inputting an SAR image to be segmented, and calculating a segmentation characteristic of the image at pixels; secondly optionally extracting a small quantity of pixels from the SAR image as a sample set S; thirdly, calculating similarity of any two pixels in the sample set, and calculating weight of sample pixels according to the similarity; fourthly, clustering the pixels in the sample set by polychotomy weighting segmentation; fifthly, calculating calculating similarity between the pixels not sampled and the sample pixels, classifying the pixels not sampled according to a neighbor principle to acquire initial segmentation of the image; and sixthly regulating segmentation of the pixels according to vote result of representative pixels by a majority principle to acquire a final segmentation result of the image. The method has the advantages of high segmentation quality and noise resistance, capability of being used for automatic target identification and the like.

Description

Based on the SAR image segmentation method of many points of Weighted Cuts

Technical field

The invention belongs to technical field of image processing, relate generally to SAR image segmentation, can be used for the object detection and recognition of SAR image.

Background technology

SAR (Synthetic Aperture Radar) is a kind of imaging system being operated in microwave region, due to its have round-the-clock, round-the-clock, from various visual angles, multiresolution data retrieval capabilities, therefore national defence and civilian in have important application.In recent years along with SAR application surface is more and more wider, the SAR data obtained from various channel gets more and more, traditional artificial interpretation be difficult to real-time from the SAR data of gained, obtain required information, utilize Target Recognition to become current study hotspot to the precision of the process and the identification of raising target of accelerating data, SAR image segmentation is then basis and the prerequisite of SAR image Motion parameters.At present many SAR image segmentation method are proposed, as the dividing method based on threshold value, based on the dividing method at edge, based on the method etc. that region increases.Although these methods all achieve good effect in the SAR image that some are concrete, the part but they still come with some shortcomings, such as when two classification target physical characteristicss comparatively close to time, based on the problem that the method for threshold value will face threshold value and is difficult to choose, inappropriate threshold value is divided causing the mistake of SAR image, divides by mistake.Particularly, the coherent speckle noise that SAR image is intrinsic also has important impact to the selection of threshold value.Also there is similar problem in other SAR image segmentation method.

In recent years, a focus in Iamge Segmentation research is become based on the image partition method of figure.First the method constructs a weighted graph when carrying out Iamge Segmentation based on the pixel of image to be split, then by carrying out figure splitting the segmentation result that can obtain image.Jianbo Shi etc. proposes Normalized cut parted pattern in " Normalized cuts and imagesegmentation ".This model is in the similarity considering to also contemplate while similarity between class in class.But when carrying out Iamge Segmentation, second the minimal characteristic vector that only make use of normalization Laplacian matrix due to Normalized cut carries out cluster, do not consider other clustering information contained in collection of illustrative plates, so the effect of Iamge Segmentation is often unsatisfactory.As improvement, Andrew Y.Ng etc. propose to utilize the clustering information in collection of illustrative plates contained by multiple proper vector to realize the segmentation of image in " Onspectral clustering:analysis and an algorithm ", and give concrete algorithm.But, proper vector and the eigenwert of a calculating matrix constructed based on image slices vegetarian refreshments is needed when utilizing the method to carry out Iamge Segmentation, due to when the dimension of image is larger, the dimension of corresponding matrix is also larger, the calculating of collection of illustrative plates also will be very difficult, even unfeasible, institute is limited by very large in this way in actual applications.For this problem, JitendraMalik etc. are at " Spectral grouping using the method " in propose utilize the method for sampling solve collection of illustrative plates be difficult to calculate problem.But the method relies on overweight to sampled point, and effect is unsatisfactory sometimes.In addition, when extracting corresponding clustering information from multiple proper vectors of figure, the suggestion such as Jitendra Malik utilizes k-means method, but the effect of the method depends on hypothesis: will distant convex class be formed when the pixel of different target is mapped to the feature space of collection of illustrative plates formation in image.Because this hypothesis may not be set up in actual applications, so the segmentation effect of image is often difficult to expect.In order to improve the segmentation effect of image further, Tian Zheng etc. are in " Iamge Segmentation based on Weighted Cut " (electronic letters, vol, 2008, Vol.36 (1)) in propose weighted cut model, its maximum feature distinguishes pixel by the mode of weighting.But similar to Normalized cut method, the method also only make use of the clustering information in second minimal characteristic vector of figure matrix, do not consider other clustering information contained in collection of illustrative plates, this makes the segmentation effect of image receive impact.

Summary of the invention

The object of the invention is to the deficiency for above-mentioned prior art, propose a kind of SAR image segmentation method based on many points of Weighted Cuts, improve the segmentation effect of SAR image.

The technical scheme realizing the object of the invention is: to utilizing in multiple proper vectors of corresponding figure matrix contained clustering information to carry out cluster to the pixel of SAR image while of pixel weighting, the method simultaneously adopting iteration to upgrade extracts clustering information from multiple proper vectors of figure, adopt the cluster result of method to pixel representing pixel ballot to adjust, concrete steps comprise as follows:

(1) input SAR image to be split, formed pixel set with all pixels in image for element, be designated as I={p ₁, p ₂..., p _n, wherein p _λrepresent λ pixel, λ=1,2 ..., n, n represent the number of pixel in SAR image;

(2) the pixel p in pixel set I is calculated _λsegmentation feature F _λ, the segmentation feature composition segmentation characteristic set F={F of all pixels ₁, F ₂..., F _n;

(3) from pixel set I, m pixel is randomly drawed, the set of composition sampled pixel point wherein represent i-th pixel extracted, i=1,2 ..., m, the pixel be not extracted forms the set of non-sample pixel wherein represent h the pixel be not extracted, h=1,2 ..., n-m;

(4) to any two pixels in sampled pixel point S set with utilize image in the segmentation feature at these two pixel places with calculate their similarity i, j=1,2 ..., m, and with for element forms the matrix on m × m rank

(5) to any one pixel in sampled pixel point S set utilize matrix A ^*computed image is at the weights of this point i=1,2 ..., m;

(6) utilize many points of Weighted Cuts to carry out cluster to the pixel in sampled pixel point S set, obtain k class C _μ, μ=1,2 ..., k;

(7) to any one pixel in non-sample pixel set U with any one pixel in sampled pixel point S set utilize image in the segmentation feature at these two pixel places with calculate their similarity h=1,2 ..., n-m, i=1,2 ..., m;

(8) to the pixel in non-sample pixel set U according to neighbour's principle, it is sorted out, obtain the initial segmentation result of image;

(9) initial segmentation result of image is adjusted according to the voting results representing pixel according to majority principle, namely to any one pixel p in pixel set I _λ, according to representing the class label of pixel to pixel p _λclass label vote, and according to majority principle to pixel p _λclass label adjust, obtain the final segmentation result of image.

The present invention compared with prior art, has following features and effect:

1) distinguishing pixel the mode both by weighting during Image Segmentation Using, utilize again clustering information contained in multiple proper vectors of figure matrix to sort out pixel simultaneously;

2) according to the relation between many points of Weighted Cut optimum solutions and its approximate solution, adopt the method for iteration from the multiple proper vector of figure matrix, find the approximate optimal solution of many points of Weighted Cuts;

3) to the image segmentation result obtained by neighbour's principle, majority principle is utilized to adjust segmentation result according to the voting results representing pixel.

Above-mentioned three features make the present invention can significantly improve the segmentation effect of SAR image, and this point obtains confirmation in actual tests.

Accompanying drawing explanation

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

Fig. 2 is the segmentation result schematic diagram that the present invention is applied in a secondary SAR image;

Fig. 3 is that the present invention and existing method are to the segmentation result comparison diagram of two secondary SAR image.

Embodiment

With reference to Fig. 1, specific embodiment of the invention process is as follows:

Step 1. inputs SAR image to be split, is formed pixel set, be designated as I={p with all pixels in image for element ₁, p ₂..., p _n, wherein p _λrepresent λ pixel, λ=1,2 ..., n, n represent the number of pixel in SAR image.

Step 2. calculates the pixel p in pixel set I _λsegmentation feature F _λ.

(2a) gray level of SAR image is quantized, be quantified as G level;

(2b) to pixel p _λ, in the picture with p _λcentered by extract the image block that size is w × w;

(2c) calculate with p _λcentered by the grey level histogram of w × w image block, and using its normalization as pixel p _λsegmentation feature F _λ;

(2d) repeat above-mentioned steps (2b) and (2c), obtain the segmentation feature of each pixel in pixel set I.

Step 3. randomly draws m pixel from pixel set I, the set of composition sampled pixel point wherein represent i-th pixel extracted, i=1,2 ..., m, the pixel be not extracted forms the set of non-sample pixel wherein represent h the pixel be not extracted, h=1,2 ..., n-m.

Any two pixels in step 4. pair sampled pixel point S set with utilize image in the segmentation feature at these two pixel places with calculate their similarity i, j=1,2 ..., m, and with for element forms the matrix on m × m rank

This step can adopt multiple method when implementing, such as, directly utilize the gray-scale value of pixel to calculate the similarity between pixel, utilize similarity between image wavelet transform coefficient calculations pixel etc., the present invention adopts the gray distribution features of image near pixel to calculate the similarity between pixel, and concrete steps are as follows:

(4a) any two pixels in sampled pixel point S set are calculated with between similarity

$A_{\mathrm{ij}}^{*}=\exp (-\frac{{\mathrm{\χ}}_{\mathrm{ij}}^2}{\mathrm{\σ}}),$

Wherein, σ is scale parameter, ${\mathrm{\χ}}_{\mathrm{ij}}^2=\frac{1}{2}\underset{t=1}{\overset{G}{\mathrm{\Σ}}}\frac{{(F_i^{*}\left(t\right)-F_j^{*}\left(t\right))}^2}{F_i^{*}\left(t\right)+F_j^{*}\left(t\right)},$ t () represents pixel segmentation feature value when gray level is t, represent pixel segmentation feature value when gray level is t, if (t)=0, then make (t)= (t)+0.001,

(4b) repeat step (4a), calculate the similarity between any two pixels in S, and with it for element is formed the matrix A on m × m rank ^*, namely

$A^{*}=\left(\begin{array}{cccc}{A}_{11}^{*}& A_{12}^{*}& ...& A_{1m}^{*}\\ A_{21}^{*}& A_{22}^{*}& ...& A_{2m}^{*}\\ ...& ...& ...& ...\\ A_{m1}^{*}& A_{m2}^{*}& ...& A_{\mathrm{mn}}^{*}\end{array}\right).$

Any one pixel in step 5. pair sampled pixel point S set utilize matrix A ^*computed image is at the weights of this point i=1,2 ..., m, concrete steps are as follows:

(5a) to the pixel in sampled pixel point S set computed image at the weights of this point,

$w_i^{*}={\left(\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{A}_{\mathrm{ij}}^{*}\right)}^{\frac{1}{2}};$

(5b) repeat step (5a), obtain the weights at image each pixel place in S.

Step 6. utilizes many points of Weighted Cuts to carry out cluster to the pixel in sampled pixel point S set, obtains k class C _μ, μ=1,2 ..., k, concrete steps are as follows:

(6a) set class number as k, order $w^{*}=\mathrm{diag}(w_1^{*},w_2^{*},...,w_m^{*}),$ $A_1^{*}=w^{*}{A}^{*}{w}^{*},$ $D_1^{*}=\mathrm{diag}\left(A_1^{*}e\right),$ Wherein represent with be m × m rank diagonal matrix of main diagonal element, represent with be m × m rank diagonal matrix of main diagonal element, e to be element be entirely 1 m dimensional vector, calculate front k+1 unit orthogonal minimal characteristic vector;

(6b) with front 2nd to kth+1 unit orthogonal minimal characteristic vector be Column vector groups vector matrix Y in column;

(6c) element maximum for numerical value in each row element of column vector matrix Y is set to 1, all the other elements are set to-1, obtain a new matrix, are called class label matrix X;

(6d) by the transposition Y of column vector matrix Y ^tbe multiplied with a class label matrix X left side and obtain transition matrix B;

(6e) be multiplied with a transition matrix B left side by column vector matrix Y, and element maximum for numerical value in each row element of the matrix of gained is set to 1, all the other elements are set to-1, obtain another new matrix, are given class label matrix X;

(6f) with the ω row x of class label matrix X _ωbe the class label vector of ω class, ω=1,2 ..., k, calculates the value of many minutes Weighted Cuts of sampled pixel point S set

$\frac{1}{4}\underset{\mathrm{\ω}=1}{\overset{k}{\mathrm{\Σ}}}{x}_{\mathrm{\ω}}^T(D_1-A_1)x_{\mathrm{\ω}};$

(6g) step (6d) is repeated, (6e), (6f), if the value of the many points of Weighted Cuts of the S calculated current is more than or equal to the value of the many points of Weighted Cuts of the S that the last time calculates, then stop iteration, the column vector of the class label matrix X that the last time obtains is the class label vector of each class in sampled pixel point S set;

(6h) utilize the column vector of class label matrix X to carry out cluster to the pixel in sampled pixel point S set, obtain k class C _μ, μ=1,2 ..., k.

Any one pixel in step 7. couple non-sample pixel set U with any one pixel in sampled pixel point S set utilize image in the segmentation feature at these two pixel places with calculate their similarity h=1,2 ..., n-m, i=1,2 ..., m.

This step can adopt multiple method when implementing, such as, directly utilize the gray-scale value of pixel to calculate the similarity between pixel, utilize similarity between image wavelet transform coefficient calculations pixel etc., the present invention adopts the gray distribution features of image near pixel to calculate the similarity between pixel, and concrete steps are as follows:

(7a) to the pixel in non-sample pixel set U with pixel in sampled pixel point S set be calculated as follows similarity therebetween

${\stackrel{\‾}{A}}_{\mathrm{hi}}^{*}=\exp (-\frac{{\widehat{\mathrm{\χ}}}_{\mathrm{hi}}^2}{{\mathrm{\σ}}_1}),$

Wherein, σ ₁for scale parameter, represent pixel segmentation feature value when gray level is ζ, (ζ) pixel is represented segmentation feature value when gray level is ζ, if (ζ)=0, then make (ζ)= (ζ)+0.001;

(7b) step (7a) is repeated, calculate the similarity between any one pixel in non-sample pixel set U and any one pixel in sampled pixel point S set, and with it for element is formed the matrix on (n-m) × m rank namely

${\stackrel{\‾}{A}}^{*}=\left(\begin{array}{cccc}{\stackrel{\‾}{A}}_{11}^{*}& {\stackrel{\‾}{A}}_{12}^{*}& ...& {\stackrel{\‾}{A}}_{1m}^{*}\\ {\stackrel{\‾}{A}}_{21}^{*}& {\stackrel{\‾}{A}}_{22}^{*}& ...& {\stackrel{\‾}{A}}_{2m}^{*}\\ ...& ...& ...& ...\\ {\stackrel{\‾}{A}}_{n-m,1}^{*}& {\stackrel{\‾}{A}}_{n-m,2}^{*}& ...& {\stackrel{\‾}{A}}_{n-m,m}^{*}\end{array}\right),$

Wherein represent pixel with pixel between similarity.

Pixel in step 8. couple non-sample pixel set U sort out it according to neighbour's principle, obtain the initial segmentation result of image, concrete steps are as follows:

(8a) capture vegetarian refreshments in U and in set the element that middle evaluation is maximum;

If (8b) the maximum element of middle numerical value is and pixel belong to C _μ, then by pixel class label be set to μ, if in have the value of multiple element all to equal the maximal value of this set, then appoint and get one of them element conduct

(8c) repeat step (8a) and (8b), obtain the class label of all pixels in non-sample pixel set U, thus obtain the initial segmentation result of image.

The initial segmentation result of step 9. pair image adjusts according to the voting results representing pixel according to majority principle, namely to any one pixel p in pixel set I _λ, according to representing the class label of pixel to pixel p _λclass label vote, and according to majority principle to pixel p _λclass label adjust, obtain the final segmentation result of image, concrete steps are as follows:

(9a) from pixel set I, randomly draw r pixel, composition represents pixel set wherein represent u the representative pixel extracted, its class label is l _u, this l _uclass tag set 1,2 ..., value in k}, u=1,2 ..., r, the pixel be not extracted composition is non-represents pixel set wherein represent θ the non-pixel that represents, θ=1,2 ..., n-r;

(9b) pixel set U is represented to non- ₁in any one pixel with represent pixel S set ₁in any one pixel be calculated as follows their similarity

${\stackrel{\‾}{Q}}_{\mathrm{\θu}}^{*}=\exp (-\frac{{\widetilde{\mathrm{\χ}}}_{\mathrm{\θu}}^2}{{\mathrm{\σ}}_2}),$

Wherein, θ=1,2 ..., n-r, u=1,2 ..., r, σ ₂pearl not exactly round in shape is scale parameter, represent pixel segmentation feature value when gray level is ψ, represent pixel segmentation feature value when gray level is ψ, if then make this step can adopt multiple method when implementing, such as, directly utilize the gray-scale value of pixel to calculate the similarity between pixel, utilize similarity between image wavelet transform coefficient calculations pixel etc., the present invention adopts the gray distribution features of image near pixel to calculate the similarity between pixel.

(9c) pixel set U is represented for non- ₁in pixel if it is set the element that middle numerical value is maximum, then will class label be set to l _a, the rest may be inferred, obtains U ₁in the class label of all pixels, if set in have the value of multiple element all to equal its maximal value, then appoint and get one of them element conduct

(9d) to representing pixel S set ₁in any one pixel with its former class label as the class label after adjustment, u=1,2 ..., r, obtains the class label after pixel adjustment in image thus;

(9e) step (9a) to (9d) N time is repeated, to each pixel p in pixel set I _λobtain N number of class label { T _{λ 1}, T _{λ 2}..., T _{λ N}, T _{λ d}∈ 1,2 ..., k}, d=1,2 ..., N, λ=1,2 ..., n, in practice, N is the bigger the better, and this example gets N>=30;

(9f) for the pixel p in pixel set I _λ, note b _{λ g}be that g class label is at { T _{λ 1}, T _{λ 2}..., T _{λ N}the middle number of times occurred, and g=1,2 ..., k, if b _{λ f}set { b _{λ 1}, b _{λ 2}..., b _{λ k}in the maximum element of numerical value, then by pixel p _λclass label be set to f, if set { b _{λ 1}, b _{λ 2}..., b _{λ k}in have the value of multiple element all to equal its maximal value, then appoint and get one of them element as b _{λ f};

(9g) (9f) is repeated to each pixel in pixel set I and obtain its class label, the final segmentation result of the image after being adjusted thus.

Effect of the present invention can be further illustrated by following experimental result:

(1) experimental technique

Method of the present invention is applied particularly to three width SAR image to check its performance.Meanwhile, as a comparison, under same experiment condition, in test also by Jitendra Malik etc. at " Spectral grouping usingthe method " propose image partition method and k-means method are applied to this three secondary SAR image.When utilizing k-means method to carry out Iamge Segmentation, in order to obtain good segmentation effect, in experiment, adopt the optimum run for 100 times as final image segmentation result.

(2) experiment content

Experiment 1, split a secondary SAR image by the inventive method, result as shown in Figure 2.Wherein:

Fig. 2 (a) to be size be 256 × 256 a secondary SAR image, it comprises two class targets;

Fig. 2 (b) shows the sampled pixel point randomly drawed in experiment;

Fig. 2 (c) shows and utilizes many points of Weighted Cuts to carry out the result of cluster to sampling pixel points;

Fig. 2 (d) shows and utilizes neighbour's principle to carry out sorting out the image segmentation result processing and obtain to non-sampled pixel according to the segmentation result of sampled pixel point;

The result that Fig. 2 (e) obtains after showing and adjusting image segmentation result according to majority principle according to the voting results representing pixel.

Observing Fig. 2 (d) visible, when adopting near neighbor method to sort out non-sampled pixel, owing to only make use of the class label of sampling pixel, in the segmentation of image, having occurred point pixel by mistake.

Comparison diagram 2 (e) and Fig. 2 (d) visible, according to the voting results representing pixel, after adjusting according to the class label of majority principle to pixel, the mistake of the two class target internal correct class label that divided pixel to be endowed completely.In addition, the segmentation of two class target intersection pixels have also been obtained adjustment.

Contrast visible by segmentation result and original image, the segmentation effect generally after adjustment is better, and two classification target borders are more level and smooth.

Experiment 2, by the inventive method and existing method and kmeans method are split two width SAR image, result as shown in Figure 3, wherein:

Fig. 3 (a) and Fig. 3 (b) shows two original SAR image;

Fig. 3 (c) and Fig. 3 (d) shows the result utilizing kmeans method to split two width SAR image;

Fig. 3 (e) and Fig. 3 (f) shows utilization the result that method is split two width SAR image;

Fig. 3 (g) and Fig. 3 (h) shows the result that the inventive method is split two width SAR image.

Contrast the segmentation result of three kinds of methods in two width SAR image, visible the inventive method achieves better segmentation effect, and method and kmeans method all fail to realize segmentation object preferably.

To sum up, the image partition method that the present invention is based on many points of Weighted Cuts is distinguished pixel by the method for weighting, from collection of illustrative plates, extract clustering information by the method for iteration, really can be improved the effect of Iamge Segmentation by the ballot that represents pixel according to the class label of majority principle determination pixel preferably.It is pointed out that the computing method of similarity between the pixel that adopts in the present invention can be selected according to the concrete condition of image, there is popularity and universality.

Claims (8)

1., based on a SAR image segmentation method for many points of Weighted Cuts, comprise the steps:

(1) input SAR image to be split, formed pixel set with all pixels in image for element, be designated as I={p ₁, p ₂..., p _n, wherein p _λrepresent λ pixel, λ=1,2 ..., n, n represent the number of pixel in SAR image;

(2) the pixel p in pixel set I is calculated _λsegmentation feature F _λ, the segmentation feature composition segmentation characteristic set F={F of all pixels ₁, F ₂..., F _n;

(3) from pixel set I, m pixel is randomly drawed, the set of composition sampled pixel point wherein represent i-th pixel extracted, i=1,2 ..., m, the pixel be not extracted forms the set of non-sample pixel wherein represent h the pixel be not extracted, h=1,2 ..., n-m;

(4) to any two pixels in sampled pixel point S set with utilize image in the segmentation feature at these two pixel places with calculate their similarity i, j=1,2 ..., m, and with for element forms the matrix on m × m rank $A^{*}={\left(A_{\mathrm{ij}}^{*}\right)}_{m\×m};$

(5) to any one pixel in sampled pixel point S set utilize matrix computed image is at the weights of this point i=1,2 ..., m;

(6) utilize many points of Weighted Cuts to carry out cluster to the pixel in sampled pixel point S set, obtain k class C _μ, μ=1,2 ..., k;

(7) to any one pixel in non-sample pixel set U with any one pixel in sampled pixel point S set utilize image in the segmentation feature at these two pixel places with calculate their similarity h=1,2 ..., n-m, i=1,2 ..., m;

(8) to the pixel in non-sample pixel set U according to neighbour's principle, it is sorted out, obtain the initial segmentation result of image;

(9) initial segmentation result of image is adjusted according to the voting results representing pixel according to majority principle, namely to any one pixel p in pixel set I _λ, according to representing the class label of pixel to pixel p _λclass label vote, and according to majority principle to pixel p _λclass label adjust, obtain the final segmentation result of image.

2. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, pixel p in the calculating pixel set I wherein described in step (2) _λthe Iamge Segmentation feature F at place _λ, carry out as follows:

(2a) gray level of SAR image is quantized, be quantified as G level;

(2b) to pixel p _λ, in the picture with p _λcentered by extract the image block that size is w × w;

(2c) calculate with p _λcentered by the grey level histogram of w × w image block, and using its normalization as pixel p _λsegmentation feature F _λ;

(2d) repeat above-mentioned steps (2b) and (2c), obtain the segmentation feature of each pixel in I.

3. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, wherein described in step (4) to any two pixels in sampled pixel point S set with utilize image in the segmentation feature of these two pixels with calculate their similarity carry out as follows:

(4a) any two pixels in sampled pixel point S set are calculated with between similarity

$A_{\mathrm{ij}}^{*}=\exp (-\frac{{\mathrm{\χ}}_{\mathrm{ij}}^2}{\mathrm{\σ}}),$

Wherein, σ is scale parameter, represent pixel segmentation feature F _i ^*value when gray level is t, represent pixel segmentation feature value when gray level is t, if F _i ^*(t)=0, $F_j^{*}\left(t\right)=0,$ Then make F _i ^*(t)=F _i ^*(t)+0.001, $F_j^{*}\left(t\right)=F_j^{*}\left(t\right)+0.001;$

(4b) repeat step (4a), calculate the similarity between any two pixels in S, and with it for element is formed the matrix A on m × m rank ^*, namely

$A^{*}=\left(\begin{array}{cccc}{A}_{11}^{*}& A_{12}^{*}& ...& A_{1m}^{*}\\ A_{21}^{*}& A_{22}^{*}& ...& A_{2m}^{*}\\ ...& ...& ...& ...\\ A_{m1}^{*}& A_{m2}^{*}& ...& A_{\mathrm{mm}}^{*}\end{array}\right).$

4. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, wherein described in step (5) to any one pixel in sampled pixel point S set utilize matrix A ^*computed image is at the weights of this point carry out as follows:

(5a) to the pixel in sampled pixel point S set computed image at the weights of this point,

$w_i^{*}={\left(\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{A}_{\mathrm{ij}}^{*}\right)}^{\frac{1}{2}};$

(5b) repeat step (5a), obtain the weights at image each pixel place in S.

5. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, the many points of Weighted Cuts that utilize wherein described in step (6) carry out cluster to sampled pixel point S set, carry out as follows:

(6a) set class number as k, order $w^{*}=\mathrm{diag}(w_1^{*},w_2^{*},\·\·\·,w_m^{*}),A_1^{*}=w^{*}{A}^{*}{w}^{*},D_1^{*}=\mathrm{diag}\left(A_1^{*}e\right),$ Wherein represent with be m × m rank diagonal matrix of main diagonal element, represent with be m × m rank diagonal matrix of main diagonal element, e to be element be entirely 1 m dimensional vector, calculate front k+1 unit orthogonal minimal characteristic vector;

(6b) with front 2nd to kth+1 unit orthogonal minimal characteristic vector be Column vector groups vector matrix Y in column;

(6c) element maximum for each row element intermediate value of column vector matrix Y is set to 1, all the other elements are set to-1, obtain a new matrix, are called class label matrix X;

(6d) by the transposition Y of column vector matrix Y ^tbe multiplied with a class label matrix X left side and obtain transition matrix B;

(6e) be multiplied with a transition matrix B left side by column vector matrix Y, and element maximum for each row element intermediate value of the matrix of gained is set to 1, all the other elements are set to-1, obtain another new matrix, are given class label matrix X;

(6f) with the ω row x of class label matrix X _ωbe the class label vector of ω class, ω=1,2 ..., k, calculates the value of many minutes Weighted Cuts of sampled pixel point S set

$\frac{1}{4}\underset{\mathrm{\ω}=1}{\overset{k}{\mathrm{\Σ}}}{x}_{\mathrm{\ω}}^T(D_1-A_1)x_{\mathrm{\ω}};$

(6g) step (6d) is repeated, (6e), (6f), if the value of the many points of Weighted Cuts of the S calculated current is more than or equal to the value of the many points of Weighted Cuts of the S that the last time calculates, then stop iteration, the column vector of the class label matrix X that the last time obtains is the class label vector of each class in sampled pixel point S set;

(6h) utilize the column vector of class label matrix X to carry out cluster to the pixel in sampled pixel point S set, obtain k class C _μ, μ=1,2 ..., k.

6. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, utilizes image at pixel wherein described in step (7) segmentation feature calculate with any one pixel in sampled pixel point S set similarity carry out as follows:

(7a) to the pixel in non-sample pixel set U with pixel in sampled pixel point S set be calculated as follows similarity therebetween

${\stackrel{\‾}{A}}_{\mathrm{hi}}^{*}=\exp (-\frac{{\widehat{\mathrm{\χ}}}_{\mathrm{hi}}^2}{{\mathrm{\σ}}_1}),$

Wherein, σ ₁for scale parameter, represent pixel segmentation feature value when gray level is ζ, F _i ^*(ζ) pixel is represented segmentation feature F _i ^*value when gray level is ζ, if f _i ^*(ζ)=0, then make f _i ^*(ζ)=F _i ^*(ζ)+0.001;

(7b) step (7a) is repeated, calculate the similarity between any one pixel in non-sample pixel set U and any one pixel in sampled pixel point S set, and with it for element is formed the matrix on (n-m) × m rank namely

${\stackrel{\‾}{A}}^{*}=\left(\begin{array}{cccc}{\stackrel{\‾}{A}}_{11}^{*}& {\stackrel{\‾}{A}}_{12}^{*}& ...& {\stackrel{\‾}{A}}_{1m}^{*}\\ {\stackrel{\‾}{A}}_{21}^{*}& {\stackrel{\‾}{A}}_{22}^{*}& ...& {\stackrel{\‾}{A}}_{2m}^{*}\\ ...& ...& ...& ...\\ {\stackrel{\‾}{A}}_{n-m,1}^{*}& {\stackrel{\‾}{A}}_{n-m,2}^{*}& ...& {\stackrel{\‾}{A}}_{n-m,m}^{*}\end{array}\right),$

Wherein represent pixel with pixel between similarity.

7. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, wherein described in step (8) to any one pixel in non-sample pixel set U according to neighbour's principle, it is sorted out, carries out as follows:

(8a) capture vegetarian refreshments in non-sample pixel set U and in set the element that middle evaluation is maximum;

If (8b) gather the maximum element of middle numerical value is and pixel belong to C _μ, then by pixel class label be set to μ, if in have the value of multiple element all to equal the maximal value of this set, then appoint and get one of them element conduct

(8c) repeat step (8a) and (8b), obtain the class label of all pixels in non-sample pixel set U, thus obtain the initial segmentation result of image.

8. the SAR image segmentation method based on many points of Weighted Cuts according to claim 1, wherein described in step (9) to any one pixel p in pixel set I _λ, according to representing the class label of pixel to p _λclass label vote, and according to majority principle to p _λclass label adjust, obtain the final segmentation result of image, carry out as follows:

(9a) randomly draw r pixel, composition represents pixel set wherein represent u the representative pixel extracted, its class label is l _u, this l _uclass tag set 1,2 ..., value in k}, u=1,2 ..., r, the pixel be not extracted composition is non-represents pixel set wherein represent θ the non-pixel that represents, θ=1,2 ..., n-r;

(9b) pixel set U is represented to non- ₁in any one pixel with represent pixel S set ₁in any one pixel be calculated as follows their similarity

${\stackrel{\‾}{Q}}_{\mathrm{\θu}}^{*}=\exp (-\frac{{\widetilde{\mathrm{\χ}}}_{\mathrm{\θu}}^2}{{\mathrm{\σ}}_2}),$

Wherein, θ=1,2 ..., n-r, u=1,2 ..., r, σ ₂for scale parameter, represent pixel segmentation feature value when gray level is ψ, represent pixel segmentation feature value when gray level is ψ, if ${\stackrel{\‾}{F}}_{\mathrm{\θ}}\left(\mathrm{\ψ}\right)=0,F_u^{*}\left(\mathrm{\ψ}\right)=0,$ Then make ${\stackrel{\‾}{F}}_{\mathrm{\θ}}\left(\mathrm{\ψ}\right)={\stackrel{\‾}{F}}_{\mathrm{\θ}}\left(\mathrm{\ψ}\right)+0.001,$ $F_u^{*}\left(\mathrm{\ψ}\right)=F_u^{*}\left(\mathrm{\ψ}\right)+0.001;$

(9c) pixel set U is represented for non- ₁in pixel if it is set the element that middle numerical value is maximum, then will class label be set to l _a, the rest may be inferred, obtains U ₁in the class label of all pixels, if set in have the value of multiple element all to equal its maximal value, then appoint and get one of them element conduct

(9d) to representing pixel S set ₁in any one pixel with its former class label as the class label after adjustment, u=1,2 ..., r, obtains the class label after pixel adjustment in image thus;

(9e) repeat step (9a) to (9d) N time, obtain each pixel p in pixel set I _λn number of class label { T _{λ 1}, T _{λ 2}..., T _{λ N}, T _{λ d}∈ 1,2 ..., k}, d=1,2 ..., N, λ=1,2 ..., n, N>=30;

(9f) for the pixel p in pixel set I _λ, note b _{λ g}be that g class label is at { T _{λ 1}, T _{λ 2}..., T _{λ N}the middle number of times occurred, and g=1,2 ..., k, if b _{λ f}set { b _{λ 1}, b _{λ 2}..., b _{λ k}in the maximum element of numerical value, then by pixel p _λclass label be set to f, if set { b _{λ 1}, b _{λ 2}..., b _{λ k}in have the value of multiple element all to equal its maximal value, then appoint and get one of them element as b _{λ f};

(9g) step (9f) is repeated to each pixel in pixel set I and obtain its class label, the final segmentation result of the image after being adjusted thus.