CN102903102A - Non-local-based triple Markov random field synthetic aperture radar (SAR) image segmentation method - Google Patents

Abstract

The invention discloses a non-local-based triple Markov random field synthetic aperture radar (SAR) image segmentation method and belongs to the technical field of image processing. The problems that a traditional triplet Markov field (TMF) method which is used in SAR image segmentation is poor in regional consistency and disorder in edge are solved. The method comprises steps of (1) inputting an image to be segmented; (2) initializing all pixel class marks by using fuzzy C-means (FCM) clustering; (3) initializing all pixel scene categories by using k-means and conducting iteration for scene categories by using non-local redundant information; (4) calculating potential energy of the image; (5) constructing triple Markov random field joint distribution, conducting function sampling for the distribution by using a Gibbs sampler and obtaining the posterior probability; (6) calculating the edge posterior probability and updating all pixel class marks gradually; and (7) determining whether the change rate of all pixel class marks is larger than the threshold, repeating step (4), step (5) and step (6) if the change rate of all pixel class marks is larger than the threshold, and inputting segmentation results if the change rate of all pixel class marks is not larger than the threshold. The method has the advantages of being quick in convergence velocity, good in segmentation result regional consistency, capable of retaining complete information and applied to SAR image target identification.

Description

Based on non-local three Markov random field SAR image partition methods

Technical field

The invention belongs to image processing field, relate to a kind of method to the inhomogeneous non-stationary SAR image segmentation of grain distribution, can be applicable to target identification.

Background technology

Synthetic-aperture radar (Synthetic Aperture Radar, SAR) is a kind of high-resolution radar system, can be applicable to the numerous areas such as military affairs, agricultural, navigation, geographical supervision.It and other remote-sensing imaging system, optical imaging system has been compared a lot of differences.Aspect military target identification, the SAR image is widely used in object detection field, and the SAR image segmentation then is the important step of processing graphical analysis from image, is the basis of target classification and identification.

Aspect the processing of SAR image segmentation; because the existence of the intrinsic property taken advantage of speckle noise in the SAR image; the pixel of image usually has sudden change; this sudden change is partial isolated; and the local correlations let us of image has considered that center pixel is adjacent the correlativity between the pixel; the Markov random field (Markov Random Field, MRF) that is based upon in the neighborhood of pixels system just in time can utilize neighborhood relevance to eliminate the impact of noise, is suitable for cutting apart of SAR image.

Since the sixties in last century, Besag etc. contact so that distribution function and energy function have had effectively about the research of gibbs (Gibbs) distribution with the MRF equivalence, utilize this contact MRF model to begin to be applied at the image processing method face.Geman S. and Geman D. have set up the method based on Markov random field and maximum a posteriori probability in the famous paper of delivering in 1984.MRF has in processing and has obtained gratifying result aspect the image of simple textures, but for the large amount of complex image, processing such as non-stationary image, complex texture image, strong noise image, it is too simple that the MRF model seems, because there are two random fields in hypothesis in the MRF model, markov distributes and only one of them random field is thought to meet, like this when processing complicated image, it is too simple that the hypothesis of the independence that can't satisfy condition, model just seem.Along with the continuous development and improvement of statistical model, the MRF model is thereupon constantly development also.2000, a double Markov random fields model (Pairwise Markov random Fields has been proposed, PMF), the joint distribution of directly supposing two random fields in this model meets the markov distribution, with regard to so that we can carry out modeling to the texture features of different images on the one hand, can utilize again on the other hand different bayes methods to realize cutting apart like this.People have arrived this model extension again three more general Markov random fields (Triplet Markov Fields, TMF) model on this basis, come image is carried out modeling by introducing the 3rd random field in this model.

Non-local mean is usually used in image denoising, is a popularization to bilateral filtering, often includes many redundant informations in the image, takes full advantage of these redundant informations for removing the picture noise service, and this is the main thought of non-local average Filtering Model.Redundant information refers to the similarity degree of subregion gray scale in the image, carries out the advantage that smoothing denoising is the non-local mean image denoising according to similarity.The principal feature of non-local mean (NL-means) model is: the method is not that the gray-scale value with single pixel in the image compares, but the distribution situation of the whole gray scale around this pixel is compared, contribute weights according to the similarity of intensity profile.

Increasingly mature along with SAR image partition method under the statistical model, the evaluation of segmentation effect are also harsh gradually aspect following three: the internal consistency of homogeneous region; The integrality of detailed information; The clarity of [Dan.Traditional TMF method is not considered the similarity of image self, therefore the contextual information of image is not fully utilized, so that in cutting procedure, lost detailed information and the edge of some images, cause this TMF method can produce mistake to the SAR image that comprises complex texture and cut apart and the phenomenons such as regional consistance is undesirable.

Summary of the invention

The object of the invention is to the deficiency for above-mentioned prior art, propose a kind of based on non-local three Markov random field SAR image partition methods, utilize the partial structurtes information similarity of image, improve edge resolving accuracy and regional consistance, guarantee the image segmentation information integrity, improve the quality of SAR image segmentation.

The technical thought that realizes the object of the invention is: adopt fuzzy C-mean algorithm (FCM) clustering method to obtain the initial classes mark field X of image, with the initial scene category label field U of the method for K mean cluster, utilize the non local redundant information of image that the U field is processed, and by TMF model description image information, in conjunction with the Bayesian theory X is upgraded, satisfy accuracy requirement or the maximum iteration time of having set until obtain class mark field X, export final segmentation result.

Its specific implementation step is as follows:

1) input SAR image to be split;

2) utilize the method for FCM cluster to obtain the initial classes mark of each pixel of image to be split;

3) gray level co-occurrence matrixes of extraction image to be split, adopt the k-means clustering method to obtain the scene classification of each pixel of image to be split, and utilize the non local redundant information of image that the scene classification is carried out an iteration and upgrade, obtain the new scene classification of each pixel;

4) according to step 2) initial classes mark and the step 3 of each pixel of obtaining) the new scene classification of each pixel that obtains, utilize following formula to calculate potential energy W (x, u):

$W(x,u)=\underset{(s,t)\∈C_H}{\mathrm{\Σ}}{\mathrm{\α}}_H^1(1-2\mathrm{\δ}(x_s,x_t))-({\mathrm{\α}}_{\mathrm{aH}}^2{\mathrm{\δ}}^{*}(u_s,u_t,a)+{\mathrm{\α}}_{\mathrm{bH}}^2{\mathrm{\δ}}^{*}(u_s,u_t,b))(1-\mathrm{\δ}(x_s,x_t))$

$+\underset{(s,t)\∈C_V}{\mathrm{\Σ}}{\mathrm{\α}}_V^1(1-2\mathrm{\δ}(x_s,x_t))-({\mathrm{\α}}_{\mathrm{aV}}^2{\mathrm{\δ}}^{*}(u_s,u_t,a)+{\mathrm{\α}}_{\mathrm{bV}}^2{\mathrm{\δ}}^{*}(u_s,u_t,b))(1-\mathrm{\δ}(x_s,x_t))$

In the formula, x is the class mark of pixel, and u is the scene classification of pixel, and s, t are a pair of neighbor pixel, x _s, x _tBe the class scale value of a pair of neighbor pixel, u _s, u _tBe the scene classification of a pair of neighbor pixel, C _HBe the right set of the neighbor pixel on the horizontal direction in the image, C _VBe the right set of the neighbor pixel on the vertical direction in the image, a, b are two kinds of values of scene classification,

With

Represent respectively different weight parameter, ${\mathrm{\α}}_H^1={\mathrm{\α}}_V^1=1,$ ${\mathrm{\α}}_{\mathrm{aH}}^2=1,$ ${\mathrm{\α}}_{\mathrm{aV}}^2=-0.3,$ ${\mathrm{\α}}_{\mathrm{bH}}^2=-0.3,$ ${\mathrm{\α}}_{\mathrm{bV}}^2=1,$ $\mathrm{\δ}(x_s,x_t)=\left\{\begin{array}{cc}1,& x_s=x_t\\ 0,& x_s\≠x_t\end{array}\right.$ Be class target penalty, ${\mathrm{\δ}}^{*}(u_s,u_t,a)=\left\{\begin{array}{cc}1,& u_s=u_t=a\\ 0,& \mathrm{otherwise}\end{array}\right.$ Penalty when getting a for the scene classification, ${\mathrm{\δ}}^{*}(u_s,u_t,b)=\left\{\begin{array}{cc}1,& u_s=u_t=b\\ 0,& \mathrm{otherwise}\end{array}\right.$ Penalty when getting b for the scene classification;

5) according to step 2) in the class mark of each pixel of obtaining, calculate the conditional probability of each pixel:

$p\left(y_s\right|x_s)=\frac{n^n\mathrm{\Γ}(n-\mathrm{\α})}{{\mathrm{\γ}}^{\mathrm{\α}}\mathrm{\Γ}\left(n\right)\mathrm{\Γ}(-\mathrm{\α})}\frac{y_s^{2(n-1)}}{{(\mathrm{\γ}+n{y}_s^2)}^{n-\mathrm{\α}}},$ -α，γ,n>0

Wherein, p (y _s| x _s) obedience statistical model G ⁰Distribute y _sBe the gray-scale value of pixel s, Γ is the Gamma function, and n is the equivalent number of SAR image, and n obtains by the priori of image, and α is form parameter, and γ is scale parameter;

6) according to the conditional probability p (y of potential energy W obtained above (x, u) and each pixel _s| x _s), upgrade the class mark of each pixel;

6a) according to the conditional probability p (y of potential energy W (x, u) and each pixel _s| x _s), use the statistical probability formula to calculate three Markov Random Fields joint probability distribution;

6b) utilize three Markov Random Fields joint probability distribution, obtain the posterior marginal probability of each pixel;

6c) utilize the class mark of each pixel in the Bayesian MAP marginal probability criterion pointwise update image;

7) will upgrade before and after in the pixel number that changes of class mark and the image ratio of pixel sum as end condition, if ratio is greater than the threshold value 10 of input ^-9, or reach maximum iteration time 100, return step 4), otherwise the final class mark of each pixel is exported as final segmentation result.

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

1, combining image of the present invention has this character of repetitive structure, takes full advantage of the redundant information of image, extracts the analog structure in non-stationary SAR image texture zone, has guaranteed the integrality of image information;

2, the present invention has suppressed speckle noise because non local characteristic and TMF model are combined, and has effectively kept again the smooth edge part in the image simultaneously, has improved the segmentation precision of SAR image;

3, the present invention is owing to adopting G ⁰The probability density function that distributes has well mated the data model of different scene areas as posterior marginal probability, has improved homogeney, connectedness and the fidelity of segmentation result.

Description of drawings

Fig. 1 is schematic flow sheet of the present invention;

Fig. 2 is the simulation result figure of the present invention on a width of cloth two analoglike SAR images;

Fig. 3 is the simulation result figure of the present invention on the complicated surface feature background SAR of a width of cloth two classes image;

Fig. 4 is the simulation result figure of the present invention on the complicated surface feature background SAR of secondary three classes image.

Embodiment

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

Step 1 is inputted SAR image to be split, and the image size is M*N, for n, comprises the targets such as farmland, waters, forest, cities and towns, mountain region depending on number in the image.

Step 2 utilizes the method for FCM cluster to obtain the initial classes mark of each pixel of image to be split.

SAR image to be split is carried out the FCM cluster, and each pixel belongs to all kinds of with different degrees of membership.Have the class of high degree of membership by after the iteration optimizing each pixel being distributed to this pixel, just obtained the initial classes mark of each pixel in the image.

Step 3, extract the gray level co-occurrence matrixes of image to be split, adopt the k-means clustering method to obtain the scene classification of each pixel of image to be split, and utilize the non local redundant information of image that the scene classification is carried out an iteration and upgrade, obtain the new scene classification of each pixel.

(3.1) gray level co-occurrence matrixes of extraction image to be split, adopt the k-means cluster to estimate the scene classification:

Choose 0 degree of each pixel, 45 degree, 90 degree, the contrast on the 135 degree four directions, correlativity, energy, unfavourable balance distance be totally 16 dimensional features; K pixel of random selection, wherein each pixel represents the cluster centre of a scene classification, calculate remaining each pixel to the distance based on above-mentioned 16 dimensional features at each scene classification center, according to making the minimum principle of distance that remaining each pixel is divided into the most similar scene classification; Then recomputate the new average of each scene classification, obtain new scene classification center; Constantly repeat this process, until scene classification center no longer changes, at this moment just finished the initial estimation to the scene classification;

(3.2) utilizing the non local redundant information of image that the scene classification is carried out an iteration upgrades.

The initial scene classification of each the pixel s that obtains according to step (3.1), get neighborhood window m*m, m=11, interior each the class pixel of calculation window is to the factor of influence of s respectively, scene classification after one class of getting the factor of influence maximum is upgraded as the s point, k class pixel to the computing formula of the factor of influence of s is:

$p(u_s=k)=\underset{u_t=k}{\mathrm{\Σ}}\frac{1}{Z\left(s\right)}\exp (-\frac{{\left|\right|v\left(N_s\right)-v\left(N_t\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}),$

In the formula, t is the pixel that belongs to the k class in the large window of neighborhood, v (N _s) the neighborhood vector of representative centered by s, v (N _t) the neighborhood vector of representative centered by t, get near the 3*3 wicket of central point totally nine somes composition neighborhood vectors, use gaussian kernel function

Similarity between tolerance neighborhood vector, the higher then factor of influence of similarity value is larger.Z (s) is normalized factor, and its expression formula is:

$Z\left(s\right)=\underset{t}{\mathrm{\Σ}}\exp (-\frac{{\left|\right|v\left(N_s\right)-v\left(N_t\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}),$

Wherein, h is the decay factor of the control characteristic function rate of decay, and α is the standard deviation of gaussian kernel function, α〉0.

Step 4, the initial classes mark of each pixel that obtains according to step 2 and the new scene classification of each pixel that step 3 obtains, utilize the potential energy W (x, u) of following formula computed image:

$W(x,u)=\underset{(s,t)\∈C_H}{\mathrm{\Σ}}{\mathrm{\α}}_H^1(1-2\mathrm{\δ}(x_s,x_t))-({\mathrm{\α}}_{\mathrm{aH}}^2{\mathrm{\δ}}^{*}(u_s,u_t,a)+{\mathrm{\α}}_{\mathrm{bH}}^2{\mathrm{\δ}}^{*}(u_s,u_t,b))(1-\mathrm{\δ}(x_s,x_t))$

$+\underset{(s,t)\∈C_V}{\mathrm{\Σ}}{\mathrm{\α}}_V^1(1-2\mathrm{\δ}(x_s,x_t))-({\mathrm{\α}}_{\mathrm{aV}}^2{\mathrm{\δ}}^{*}(u_s,u_t,a)+{\mathrm{\α}}_{\mathrm{bV}}^2{\mathrm{\δ}}^{*}(u_s,u_t,b))(1-\mathrm{\δ}(x_s,x_t))$

In the formula, x is the class mark of pixel, and u is the scene classification of pixel, and s, t are a pair of neighbor pixel, x _s, x _tBe the class scale value of a pair of neighbor pixel, u _s, u _tBe the scene classification of a pair of neighbor pixel, C _HBe the right set of the neighbor pixel on the horizontal direction in the image, C _VBe the right set of the neighbor pixel on the vertical direction in the image, a, b are two kinds of values of scene classification,

With

Represent respectively different weight parameter, ${\mathrm{\α}}_H^1={\mathrm{\α}}_V^1=1,$ ${\mathrm{\α}}_{\mathrm{aH}}^2=1,$ ${\mathrm{\α}}_{\mathrm{aV}}^2=-0.3,$ ${\mathrm{\α}}_{\mathrm{bH}}^2=-0.3,$ ${\mathrm{\α}}_{\mathrm{bV}}^2=1,$ $\mathrm{\δ}(x_s,x_t)=\left\{\begin{array}{cc}1,& x_s=x_t\\ 0,& x_s\≠x_t\end{array}\right.$ Be class target penalty, ${\mathrm{\δ}}^{*}(u_s,u_t,a)=\left\{\begin{array}{cc}1,& u_s=u_t=a\\ 0,& \mathrm{otherwise}\end{array}\right.$ Penalty when getting a for the scene classification, ${\mathrm{\δ}}^{*}(u_s,u_t,b)=\left\{\begin{array}{cc}1,& u_s=u_t=b\\ 0,& \mathrm{otherwise}\end{array}\right.$ Penalty when getting b for the scene classification.

Step 5, according to the class mark of each pixel that obtains in the step 2, calculate the conditional probability of each pixel:

$p\left(y_s\right|x_s)=\frac{n^n\mathrm{\Γ}(n-\mathrm{\α})}{{\mathrm{\γ}}^{\mathrm{\α}}\mathrm{\Γ}\left(n\right)\mathrm{\Γ}(-\mathrm{\α})}\frac{y_s^{2(n-1)}}{{(\mathrm{\γ}+n{y}_s^2)}^{n-\mathrm{\α}}},$

Wherein, p (y _s| x _s) obedience statistical model G ⁰Distribute y _sBe the gray-scale value of pixel s, Γ is the Gamma function, and n is the equivalent number of SAR image, and n can obtain by the priori of image, and α is form parameter, and γ is scale parameter.

Step 6 is according to the conditional probability p (y of potential energy W obtained above (x, u) and each pixel _s| x _s), upgrade the class mark of each pixel.

(6.1) according to the conditional probability p (y of potential energy W (x, u) and each pixel _s| x _s), use the statistical probability formula to calculate three Markov Random Fields joint probability distribution:

The statistical probability formula is:

$p(x,u,y)={\mathrm{\γ}}^{\′}\exp [-W(x,u)+\underset{s\∈S}{\mathrm{\Σ}}\mathrm{Log}\left(p\left(y_s\right|x_s)\right)],$

In the formula, p (x, u, y) is the three Markov Random Fields joint probability distribution of class mark x, scene classification u and gray-scale value y, and γ ' is normaliztion constant, and W (x, u) is potential energy, and s is pixel, p (y _s| x _s) be conditional probability, y _sBe the gray-scale value of pixel s, x _sClass mark for pixel s;

(6.2) utilize three Markov Random Fields joint probability distribution, obtain the posterior marginal probability of each pixel:

Joint probability distribution to three Markov Random Fields is carried out gibbs sampler, sample during 30-60 class mark set and the scene classification of sampling gathered, select the class scale value that each pixel occurrence number is maximum in the sample, utilize following formula to calculate the posterior probability p (x of each pixel _s, u _s| y):

$p(x_s,u_s|y)=\frac{p(x,u,y)}{p\left(y\right)},$

Wherein, p (y) is the probability distribution of gradation of image, and p (x, u, y) is three Markov Random Fields joint probability distribution;

Utilize following formula to calculate the posterior marginal probability p (x of each pixel _s| y):

$p\left(x_s\right|y)={\mathrm{\Σ}}_{u_s\∈^}p(x_s,u_s|y),$

In the formula, s is pixel, u _sBe the scene classification of pixel, ^ is the set under the scene classification, x _sBe the mark of pixel, y is the pixel gray-scale value, p (x _s| y) be posterior marginal probability;

(6.3) upgrade the class mark of each pixel according to Bayesian MAP marginal probability criterion:

${\hat{x}}_s=\arg \max p_{\mathrm{\ω}\∈\mathrm{\Ω}}(x_s={\mathrm{\ω}}_i|y)$

Wherein, ω _iBe the possible class scale value of s point, i=1 ..., k, Ω are the set under the class scale value, Ω={ ω ₁, ω ₂..., ω _k, k is the class mark sum of cutting apart, x _sFor choosing one by one the class mark in the set of class mark, p _{ω ∈ Ω}(x _s=ω _i| y) be the posterior marginal probability of class mark set, argmax is the maximizing symbol,

For so that p _{ω ∈ Ω}(x _s=ω _i| the class mark when y) reaching maximal value, will

As the class mark after the renewal of s point.

Step 7, the ratio of pixel sum is as end condition, if ratio is greater than the threshold value 10 of input in the pixel number that class mark before and after upgrading is changed and the image ^-9, or reach maximum iteration time 100, return step 4), otherwise the final class mark of each pixel is exported as final segmentation result.

Effect of the present invention can further specify by following emulation:

1 simulated conditions

Emulation of the present invention is to carry out under the software environment of the hardware environment of the Intel of dominant frequency 2.67GHZ (R) Core (TM) i5 CPU M480, internal memory 3.80GB and MATLAB R2010b.

The image that 2 emulation experiments adopt

Experiment uses the big or small texture SAR image as 256*256 of 3 width of cloth shown in Fig. 2 (a), Fig. 3 (a), Fig. 4 (a) as test pattern.

3. emulation content

Emulation 1 uses the inventive method and traditional TMF method to two analoglike SAR Image Segmentation Usings shown in Fig. 2 (a), and segmentation result is shown in Fig. 2 (b) and Fig. 2 (c).Wherein:

Fig. 2 (b) is for to cut apart the result who obtains with traditional TMF method to Fig. 2 (a);

Fig. 2 (c) is for to cut apart the result who obtains with the inventive method to Fig. 2 (a).

By Fig. 2 (b) as seen, cut apart background parts with traditional TMF method more in disorder, region contour discrimination ability is bad, and the misclassification blocking effect clearly, and this effect is owing to the shortcoming of TMF model to picture structure information correlativity and probability distribution description to a great extent;

By Fig. 2 (c) as seen, regional consistance of the present invention is better, and the edge clear detailed information is complete.

Emulation 2 uses the inventive method and traditional TMF method to the complicated surface feature background SAR of two classes shown in Fig. 3 (a) Image Segmentation Using.Segmentation result is shown in Fig. 3 (b) and Fig. 3 (c).Wherein:

Fig. 3 (b) is cut apart the result who obtains for traditional TMF method to Fig. 3 (a);

Fig. 3 (c) is cut apart the result who obtains for the inventive method to Fig. 3 (a).

By Fig. 3 (b) as seen, obvious burrs on edges appears in traditional TMF method, and regional consistance is poor;

By Fig. 3 (c) as seen, the present invention is accurately more clear for the reservation of detailed information, and the border is smooth continuously.

Emulation 3 uses the inventive method and traditional TMF method to the complicated surface feature background SAR of three classes shown in Fig. 4 (a) Image Segmentation Using.Segmentation result is shown in Fig. 4 (b) and Fig. 4 (c).Wherein:

Fig. 4 (b) is cut apart the result who obtains for traditional TMF method to Fig. 4 (a);

Fig. 4 (c) is cut apart the result who obtains for the inventive method to Fig. 4 (a).

By Fig. 4 (c) as seen, zone of the present invention is smoother continuously, and comparatively complete for the complicated ground object target extraction in the SAR image, such as cities and towns part among Fig. 4 (a), the present invention can be cut apart preferably.

Claims (5)

1. one kind based on non-local three Markov random field SAR image partition methods, comprise the steps:

1) input SAR image to be split;

2) utilize the method for FCM cluster to obtain the initial classes mark of each pixel of image to be split;

3) gray level co-occurrence matrixes of extraction image to be split, adopt the k-means clustering method to obtain the scene classification of each pixel of image to be split, and utilize the non local redundant information of image that the scene classification is carried out an iteration and upgrade, obtain the new scene classification of each pixel;

4) according to step 2) initial classes mark and the step 3 of each pixel of obtaining) the new scene classification of each pixel that obtains, utilize following formula to calculate potential energy W (x, u):

$W(x,u)=\underset{(s,t)\∈C_H}{\mathrm{\Σ}}{\mathrm{\α}}_H^1(1-2\mathrm{\δ}(x_s,x_t))-({\mathrm{\α}}_{\mathrm{aH}}^2{\mathrm{\δ}}^{*}(u_s,u_t,a)+{\mathrm{\α}}_{\mathrm{bH}}^2{\mathrm{\δ}}^{*}(u_s,u_t,b))(1-\mathrm{\δ}(x_s,x_t))$

$+\underset{(s,t)\∈C_V}{\mathrm{\Σ}}{\mathrm{\α}}_V^1(1-2\mathrm{\δ}(x_s,x_t))-({\mathrm{\α}}_{\mathrm{aV}}^2{\mathrm{\δ}}^{*}(u_s,u_t,a)+{\mathrm{\α}}_{\mathrm{bV}}^2{\mathrm{\δ}}^{*}(u_s,u_t,b))(1-\mathrm{\δ}(x_s,x_t))$

In the formula, x is the class mark of pixel, and u is the scene classification of pixel, and s, t are a pair of neighbor pixel, x _s, x _tBe the class scale value of a pair of neighbor pixel, u _s, u _tBe the scene classification of a pair of neighbor pixel, C _HBe the right set of the neighbor pixel on the horizontal direction in the image, C _VBe the right set of the neighbor pixel on the vertical direction in the image, a, b are two kinds of values of scene classification,

With

Represent respectively different weight parameter, ${\mathrm{\α}}_H^1={\mathrm{\α}}_V^1=1,$ ${\mathrm{\α}}_{\mathrm{aH}}^2=1,$ ${\mathrm{\α}}_{\mathrm{aV}}^2=-0.3,$ ${\mathrm{\α}}_{\mathrm{bH}}^2=-0.3,$ ${\mathrm{\α}}_{\mathrm{bV}}^2=1,$ $\mathrm{\δ}(x_s,x_t)=\left\{\begin{array}{cc}1,& x_s=x_t\\ 0,& x_s\≠x_t\end{array}\right.$ Be class target penalty, ${\mathrm{\δ}}^{*}(u_s,u_t,a)=\left\{\begin{array}{cc}1,& u_s=u_t=a\\ 0,& \mathrm{otherwise}\end{array}\right.$ Penalty when getting a for the scene classification, ${\mathrm{\δ}}^{*}(u_s,u_t,b)=\left\{\begin{array}{cc}1,& u_s=u_t=b\\ 0,& \mathrm{otherwise}\end{array}\right.$ Penalty when getting b for the scene classification;

5) according to step 2) in the class mark of each pixel of obtaining, calculate the conditional probability of each pixel:

$p\left(y_s\right|x_s)=\frac{n^n\mathrm{\Γ}(n-\mathrm{\α})}{{\mathrm{\γ}}^{\mathrm{\α}}\mathrm{\Γ}\left(n\right)\mathrm{\Γ}(-\mathrm{\α})}\frac{y_s^{2(n-1)}}{{(\mathrm{\γ}+n{y}_s^2)}^{n-\mathrm{\α}}},$

Wherein, p (y _s| x _s) obedience statistical model G ⁰Distribute y _sBe the gray-scale value of pixel s, Γ is the Gamma function, and n is the equivalent number of SAR image, and n obtains by the priori of image, and α is form parameter, and γ is scale parameter;

6) according to the conditional probability p (y of potential energy W obtained above (x, u) and each pixel _s| x _s), upgrade the class mark of each pixel;

6a) according to the conditional probability p (y of potential energy W (x, u) and each pixel _s| x _s), use the statistical probability formula to calculate three Markov Random Fields joint probability distribution;

6b) utilize three Markov Random Fields joint probability distribution, obtain the posterior marginal probability of each pixel;

6c) utilize the class mark of each pixel in the Bayesian MAP marginal probability criterion pointwise update image;

7) will upgrade before and after in the pixel number that changes of class mark and the image ratio of pixel sum as end condition, if ratio is greater than the threshold value 10 of input ^-9, or reach maximum iteration time 100, return step 4), otherwise the final class mark of each pixel is exported as final segmentation result.

2. according to claim 1 based on non-local three Markov random field SAR image partition method, wherein steps 3) describedly utilize the non local redundant information of image that the scene classification is carried out iteration to upgrade, carry out as follows:

3a) the gray level co-occurrence matrixes of extraction image to be split adopts the k-means cluster to estimate the initial scene classification of each pixel;

3b) according to step 3a) the initial scene classification of each pixel s of obtaining, get neighborhood window m*m, m=11, interior each the class pixel of calculation window is to the factor of influence of s respectively, scene classification after one class of getting the factor of influence maximum is upgraded as the s point, k class pixel to the computing formula of the factor of influence of s is:

$p(u_s=k)=\underset{u_t=k}{\mathrm{\Σ}}\frac{1}{Z\left(s\right)}\exp (-\frac{{\left|\right|v\left(N_s\right)-v\left(N_t\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}),$

In the formula, t is the pixel that belongs to the k class in the large window of neighborhood, v (N _s) the neighborhood vector of representative centered by s, v (N _t) the neighborhood vector of representative centered by t,

Be gaussian kernel function, Z (s) is normalized factor, and its expression formula is:

$Z\left(s\right)=\underset{t}{\mathrm{\Σ}}\exp (-\frac{{\left|\right|v\left(N_s\right)-v\left(N_t\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}),$

Wherein, h is the decay factor of the control characteristic function rate of decay, and α is the standard deviation of gaussian kernel function, α〉0.

3. according to claim 1 based on non-local three Markov random field SAR image partition method, wherein step 6a) described use statistical probability formula calculates three Markov Random Fields joint probability distribution, undertaken by following formula:

$p(x,u,y)={\mathrm{\γ}}^{\′}\exp [-W(x,u)+\underset{s\∈S}{\mathrm{\Σ}}\mathrm{Log}\left(p\left(y_s\right|x_s)\right)],$

In the formula, p (x, u, y) is the three Markov Random Fields joint probability distribution of class mark x, scene classification u and gray-scale value y, and γ ' is normaliztion constant, and W (x, u) is potential energy, and s is pixel, p (y _s| x _s) be conditional probability, y _sBe the gray-scale value of pixel s, x _sClass mark for pixel s.

4. according to claim 1 based on non-local three Markov random field SAR image partition method, wherein step 6b) the described three Markov Random Fields joint probability distribution of utilizing, obtain the posterior marginal probability of each pixel, carry out as follows:

6b1) joint probability distribution of three Markov Random Fields is carried out gibbs sampler, sample during 30-60 class mark set and the scene classification of sampling gathered, select the class scale value that each pixel occurrence number is maximum in the sample, utilize following formula to calculate the posterior probability p (x of each pixel _s, u _s| y):

$p(x_s,u_s|y)=\frac{p(x,u,y)}{p\left(y\right)},$

Wherein, p (y) is the probability distribution of gradation of image, and p (x, u, y) is three Markov Random Fields joint probability distribution; 6b2) utilize following formula to calculate the posterior marginal probability p (x of each pixel _s| y):

$p\left(x_s\right|y)={\mathrm{\Σ}}_{u_s\∈^}p(x_s,u_s|y),$

In the formula, s is pixel, u _sBe the scene classification of pixel, ^ is the set under the scene classification, x _sBe the mark of pixel, y is the pixel gray-scale value, p (x _s| y) be posterior marginal probability.

5. according to claim 1 based on non-local three Markov random field SAR image partition method, wherein step 6c) the described class mark that utilizes each pixel in the Bayesian MAP marginal probability criterion pointwise update image, undertaken by following formula:

${\hat{x}}_s=\arg \max p_{\mathrm{\ω}\∈\mathrm{\Ω}}(x_s={\mathrm{\ω}}_i|y),$

Wherein, ω _iBe the possible class scale value of s point, i=1 ..., k, Ω are the set under the class scale value, Ω={ ω ₁, ω ₂..., ω _k, k is the class mark sum of cutting apart, x _sFor choosing one by one the class mark in the set of class mark, y is the pixel gray-scale value, p _{ω ∈ Ω}(x _s=ω _i| y) be the posterior marginal probability of class mark set, argmax is the maximizing symbol,

For so that p _{ω ∈ Ω}(x _s=ω _i| the class mark when y) reaching maximal value, will As the class mark after the renewal of s point.

Images