CN103955913A - SAR image segmentation method based on line segment co-occurrence matrix characteristics and regional maps - Google Patents

Abstract

The invention discloses an SAR image segmentation method based on line segment gray-level co-occurrence characteristics and regional maps. The method mainly includes the following steps: according to an initial sketch model, extracting a sketch of an SAR image and according to a line segment gray-level co-occurrence matrix, classifying line segments into two types: W1 and W2; according to the line segment classification result and neighbor clustering analysis of a line segment space, extracting regional maps of the SAR image; according to the regional maps, mapping the SAR image into a clustering area, a non-clustering area and a line-segment-free area; using a water-shed method to perform over-segmentation on the SAR image; adopting different combination strategies to combine the three areas of the SAR image and the integrating the combination results of the three areas; and using AP clustering to carry out class label determination on the divided areas so as to obtain an SAR image segmentation result finally. The SAR image segmentation method based on the line segment gray-level co-occurrence characteristics and the regional maps is capable of effectively solving a problem that because of an SAR special imaging mechanism, ground feature clustering areas show lightness and gray-level changes with statistical regularities on the SAR image so that areas of the same type are classified into a plurality of types.

Description

A kind of SAR image partition method based on line segment co-occurrence matrix feature and areal map

Technical field

The invention belongs to technical field of image processing, relate to the dividing method of SAR image, specifically the SAR image partition method based on line segment co-occurrence matrix feature and areal map.

Background technology

Image is cut apart to refer to according to features such as gray scale, color, texture and shapes image is divided into some mutually regions of crossover not, and makes these features in the same area, present similarity, and between zones of different, presents obvious otherness; From broadly, be exactly by the object in image, according to gray scale or other features, be divided into the process little, that be connected to each other and do not want the region of handing over that covers all image pixels.It is the most important technology of fundamental sum of carrying out graphical analysis, image understanding and iamge description that image is cut apart, usually used as the first step in target detection, classification and recognizer, the quality that image is cut apart quality directly has influence on the quality of follow-up analysis, identification etc.Also ununified evaluation judgment criterion of the quality of segmentation result is generally according to actual application scenarios, segmentation effect to be judged.But scholars' trial and research for many years, still accumulated the image partition method of a lot of classics, be broadly divided into four kinds: the image based on threshold value cuts apart, image based on edge is cut apart, image based on region characteristic is cut apart and image based on statistical pattern classification is cut apart, and SAR image to cut apart be the committed step of SAR image interpretation and understanding, the quality of cutting apart directly affects the result of SAR image processing.Due to image-forming principle and the system thereof of SAR uniqueness, determine that SAR image has a large amount of coherent speckle noises, the complicated various exclusive feature such as target, the shade mingling, this makes to become quite complicated cutting apart of SAR image.Again because SAR image and optical imagery exist very large difference, the method of many optical imagerys all can not be applied directly in SAR image, and the method for cutting apart for SAR image at present can rough segmentation be two classes: the dividing method based on gray level and the dividing method based on texture.But in actual applications, these two class methods are interdependent in the following deficiency:

(1) dividing method based on gray level, the general unit processing of these class methods is pixel or super pixel, but to due to the special imaging of SAR image, cause existing in SAR image and in class, there is the very significantly region of light and shade gray scale, there is same class region to be divided into the problem of multiclass, particularly for atural object aggregation zone in High Resolution SAR image, as groups of building or forest;

(2) dividing method based on texture, these class methods are generally to provide a description in advance model, and need sample parameter to carry out learning model parameter, although it can be partitioned into the region of region high conformity, but belonging to, these class methods have the SAR of supervision image to cut apart, automatic data processing completely, this has limited the application of these class methods.

Patented claim " based on the SAR image partition method of semantic information the classification " (publication number: CN103198479A of Xian Electronics Science and Technology University, application number: CN201310102443, the applying date: on March 27th, 2013) in a kind of SAR image partition method based on semantic information classification is disclosed, call patented method in the following text, the method is used initial sketch model to obtain the sketch map that the multiple dimensioned pixel grey scale neighborhood of presentation video changes, the aggregation numerical value and the Distributive Characters that define its line segment according to sketch map are expressed the different semantic informations that line segment contains in SAR image, and according to the semantic information of line segment, line segment is classified, based on semantic information analysis, in sketch map, obtain connected region, complete cutting apart of line segment to containing complicated grey scale change ground object area, to the remaining area in initial sketch map, realize cutting apart of this part SAR image by a kind of watershed segmentation method based on subregion, the SAR image completing based on initial sketch map and semantic information classification is cut apart.This inventive method and existing additive method have to ground object target assemble the segmentation result of the ground object area forming consistent connective better, can correctly detect all isolated feature target advantages, but the method has just been utilized the locus such as aggregation, the distributed architecture feature of initial sketch map middle conductor, the indicating characteristic that does not embody SAR image itself, this method can further be improved.

Summary of the invention

This image processing method object of the present invention is to overcome the deficiency of above-mentioned existing SAR image partition method, on theory of vision computing and initial sketch model, a kind of SAR image partition method based on line segment co-occurrence matrix feature and areal map is proposed, present regular light and shade grey scale change because the special image-forming mechanism of SAR causes atural object aggregation zone at SAR image to solve, and cause similar region to be divided into the problem of multiclass.

For achieving the above object, a kind of main concrete steps of SAR image partition method based on line segment co-occurrence matrix feature and areal map of the present invention comprise as follows:

1, according to the sketch map of initial sketch model extraction SAR image;

2, according to line segment gray level co-occurrence matrixes feature, line segment is divided into two class W1 and W2, W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces;

3, realize sketch map Spatial Semantics according to the cluster analysis of sketch line segment classification results and sketch line segment spatial neighbors and divide, the areal map that the extraction in sketch map is more abstract.

4, former SAR image mapped is become aggregation zone by this areal map, non-aggregation zone and wireless segment region.

5, use watershed algorithm to carry out prime area division to SAR image, SAR image is divided into many super pixels.

6, respectively to aggregation zone, the super pixel in non-aggregation zone and wireless segment region adopts different consolidation strategies to merge, then integrates trizonal amalgamation result, obtains the final region division result of SAR image.

7, according to the average gray of the final zoning of SAR image, use AP cluster to determine region class mark to zoning, finally obtain SAR image segmentation result.

In said method, wherein described in step 2, according to line segment gray level co-occurrence matrixes feature, line segment is divided into two class W1 and W2, W1 portrays border and line target, W2 portray light and shade grey scale change produce line segment in line segment gray level co-occurrence matrixes feature extract as follows:

(2.a) by sketch map line segment S _ithe locus at place is mapped to SAR image, then with sketch map line segment S _icentered by correspondence position, getting a size at SAR image is L _i× rectangular area (2*HS+1), wherein the length of rectangular area is line segment S _ilength L _i, direction and line segment S that rectangular area is long _idirection is parallel, and rectangle region field width is fixed value 2*HS+1, the direction of rectangle region field width and line segment S _idirection is vertical, claims Wei Xian region, this rectangular area R _i;

(2.b) by 8 intervals of 0 °～180 ° average deciles of direction, be respectively [0 ° 11.25 °) [168.75 ° of ∪, 180 °), [11.25 °, 33.75 °), [33.75 °, 56.25 °), [56.25 °, 78.75 °), [78.75 °, 101.25 °), [101.25 °, 123.75 °), [123.75 °, 146.25 °), [146.25 °, 168.75 °), and these 8 interval direction point abundances are turned to 0 ° 22.5 ° 45 ° 67.5 ° 90 ° 112.5 ° 135 ° 157.5 °, according to this quantization method, to all sketch line segment S _idirection quantize, establish sketch line segment S _iside vector turn to

(2.c) according to 2b) sketch line segment S in step _iquantized directions , get four direction (direction is mould 180 all), calculates sketch line segment S _icorresponding line region R _ifour direction gray level co-occurrence matrixes

(2.d) according to 2.c) sketch map line segment S in step _igray level co-occurrence matrixes calculate sketch map line segment S _igray level co-occurrence matrixes second-order statistic: entropy, contrast and correlativity, its computing formula is as follows:

Entropy: ${\mathrm{ENT}}_h=-\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}{G}_h(p,q)\log G_h(p,q)$

Contrast: ${\mathrm{CON}}_h=\underset{n=0}{\overset{K-1}{\mathrm{\Σ}}}{n}^2\underset{|p-q|=n}{\mathrm{\Σ}}{G}_h(p,q)$

Correlativity:

${\mathrm{COR}}_h=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}\frac{(p\·q)G_h(p,q)-u_p{u}_q}{s_p{s}_q}$

$\begin{array}{cc}{u}_p=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}p{G}_h(p,q)& u_q=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}q{G}_h(p,q)\end{array}$

$\begin{array}{cc}{S}_p^2=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}{G}_h(p,q){(p-u_p)}^2& S_q^2=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}{G}_h(p,q){(q-u_q)}^2\end{array}$

Wherein K represents the size of gray level co-occurrence matrixes, for sketch line segment S _idirection quantification amount.And according to the order of this four direction is spliced into the proper vector of one 12 dimension.

Wherein described in step 2, according to line segment gray level co-occurrence matrixes feature, line segment is divided into two class W1 and W2, W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces; Carry out as follows:

3.a) according to line segment gray level co-occurrence matrixes feature calculation method, calculate line segment gray level co-occurrence matrixes feature, utilizing this feature to use K-means clustering method is two classes by line segment cluster;

3.b) calculate sketch line segment S _ik-nearest neighbor distance a _i, computing formula is as follows:

$\begin{array}{cc}{a}_i=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{D}_{\mathrm{ij}}& D_i=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\end{array}$

Wherein (x _i, y _i), (x _j, y _j) be respectively sketch line segment S _i, S _jmiddle point coordinate, M represents and line segment S _ithe number of phase neighbour's line segment, D _ijrepresent line segment S _i, S _jeuclidean distance;

3.c) the average K-nearest neighbor distance Ma of calculating same class graticule section _i, i=1,2, computing formula is as follows:

${\mathrm{Ma}}_i=\frac{1}{n_i}\underset{j=1}{\overset{n_i}{\mathrm{\Σ}}}{a}_j,i=\mathrm{1,2}$

Wherein n _ithe number of same class graticule section.

3.d) sketch line segment little average K-nearest neighbor distance is designated as to W2 class line segment, portrays the line segment that light and shade grey scale change produces, and the average large sketch line segment of K-nearest neighbor distance is designated as W2 class line segment, portrays border and line target.

Wherein described in step 4, according to areal map, former SAR image mapped is become to aggregation zone, non-aggregation zone and wireless segment region; Carry out as follows:

4.a) according to the value of pixel in the areal map of aggregation zone, if in areal map the value of pixel be 0, in corresponding SAR image, this pixel does not belong to aggregation zone, otherwise in corresponding SAR image, this pixel belongs to aggregation zone.The value of all pixels of judging area figure, extracts the aggregation zone in SAR image.

4.b) according to the value of pixel in the areal map of non-aggregation zone, if in areal map the value of pixel be 0, in corresponding SAR image, this pixel does not belong to non-aggregation zone, otherwise in corresponding SAR image, this pixel belongs to non-aggregation zone.The value of all pixels of judging area figure, extracts the non-aggregation zone in SAR image.。

4.c) remaining area of removing aggregation zone and non-aggregation zone in SAR image is called to wireless segment region.

Wherein the described use AP cluster of step (7) is determined the class mark of zoning, finally obtains the segmentation result of SAR image, is the average gray value mG that calculates each final zoning _i, use AP cluster to determine region class mark to zoning, finally obtain SAR image segmentation result, its AP clustering parameter is set: by two region R _iand R _jthe Euclidean distance of average gray value as the similarity si in two regions _j, region R _ireference degree get region R _iwith other Regional Similarities si _j, j=1 ..., the intermediate value of h.Obtain final SAR image segmentation result.

Compared with prior art, tool has the following advantages in the present invention:

The present invention, owing to utilizing the sketch line segment of initial sketch model extraction SAR image, is divided into two class W1 and W2 according to line segment gray level co-occurrence matrixes feature line segment, and W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces; According to the statistical distribution of line segment aggregation extent, in sketch map, extract the areal map of more abstract aggregation zone; Use circular window to the sliding window of W1 class sketch line segment, in sketch map, extract the areal map of non-aggregation zone; Because the local field of line segment of definition can co-occurrence matrix have not only utilized the aggregation of sketch line segment spatial neighbors, but also utilize image statistics feature, therefore effectively effectively extract the atural object aggregation zone in SAR image, present regular light and shade grey scale change thereby solved because the special image-forming mechanism of SAR causes atural object aggregation zone at SAR image, and cause similar region to be divided into the problem of multiclass; And the present invention is respectively to aggregation zone, boundary alignment is carried out in non-aggregation zone and employing watershed divide, wireless segment region, makes to obtain on the whole the satisfied segmentation result of SAR image.Advantage is as follows:

1. show by experiment that the feature that the present invention defines is effectively, can effectively sketch line segment be divided into two class W1 and W2, W1 is the line segment of portraying border and line target, and W2 portrays the line segment that light and shade grey scale change produces.

2. show that by emulation experiment the present invention can be partitioned into the atural object aggregation zone with light and shade grey scale change of consistent connectedness, effectively solve because the special image-forming mechanism of SAR causes atural object aggregation zone and presented regular light and shade grey scale change at SAR image, and caused similar region to be divided into the problem of multiclass.And can obtain on the whole the satisfied segmentation result of SAR image.

Brief description of the drawings

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

Fig. 2 extracts areal map block diagram in sketch map in the present invention;

Fig. 3 is the former SAR image that the emulation experiment in the present invention is used;

Fig. 4 is the sketch map based on initial sketch model extraction in the present invention;

Fig. 5 is the line extracted region schematic diagram based on sketch line segment spatial positional information in the present invention;

Fig. 6 is several sketch line charts that shift to an earlier date from sketch map in the present invention;

Fig. 7 extracts the local field gray level co-occurrence matrixes characteristic statistics figure of line segment in the present invention;

Fig. 8 is sketch line segment classification results figure in the present invention;

Fig. 9 is the areal map in sketch map in the present invention;

Figure 10 is that in the present invention, areal map is aggregation zone by SAR image mapped, non-aggregation zone and wireless segment areal map;

Figure 11 be invention in dividing ridge method to SAR image over-segmentation figure;

Figure 12 is SAR image aggregation zone in invention, and the super pixel in non-aggregation zone and wireless segment region merges figure;

Figure 13 is the final area division result figure of SAR image in invention;

Figure 14 is that in invention, SAR image is finally cut apart figure.

Embodiment

Below in conjunction with embodiment accompanying drawing, the present invention is described further.

Embodiment 1, describes with reference to Fig. 1 and Fig. 2.

A SAR image partition method based on line segment co-occurrence matrix feature and areal map, concrete steps comprise as follows:

(1) according to the sketch map of initial sketch model extraction SAR image;

(2) according to line segment gray level co-occurrence matrixes feature, line segment is divided into two class W1 and W2, W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces;

(3) realize sketch map Spatial Semantics according to the cluster analysis of sketch line segment classification results and sketch line segment spatial neighbors and divide, the areal map that the extraction in sketch map is more abstract;

(4), according to areal map, former SAR image mapped is become to aggregation zone, non-aggregation zone and wireless segment region;

(5), use watershed algorithm SAR image is carried out to prime area division, SAR image is divided into many super pixels;

(6), respectively to aggregation zone, the super pixel in non-aggregation zone and wireless segment region adopts different consolidation strategies to merge, then integrates trizonal amalgamation result, obtains the final region division result of SAR image;

(7), according to the average gray of the final zoning of SAR image, use AP cluster to determine region class mark to zoning, finally obtain SAR image segmentation result.

The method has solved because the special image-forming mechanism of SAR causes atural object aggregation zone and has presented regular light and shade grey scale change at SAR image, and causes similar region to be divided into the problem of multiclass.And can obtain on the whole the satisfied segmentation result of SAR image.

Embodiment 2,1-14 describes by reference to the accompanying drawings.

On the basis of embodiment 1, described step (1), for the initial sketch model proposing according to people such as Zhu Songchun, input PIPERIVER is as shown in Figure 2 desired to make money or profit and joins together to represent with wavelet theory (sparse coding) and markov random file, extract the sketch map of SAR image, as shown in Figure 3.

Described step (2), is divided into two class W1 and W2 according to line segment gray level co-occurrence matrixes feature line segment, and W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces; Specific implementation step is as follows:

2.a) according to line segment gray level co-occurrence matrixes feature calculation method, calculate line segment gray level co-occurrence matrixes feature, calculation procedure is as follows:

2.a.1) by sketch map line segment S _ithe locus at place is mapped to SAR image, then with sketch map line segment S _icentered by correspondence position, getting a size at SAR image is L _i× rectangular area (2*HS+1), wherein the length of rectangular area is line segment S _ilength L _i, direction and line segment S that rectangular area is long _idirection is parallel, and rectangle region field width is fixed value 2*HS+1, the direction of rectangle region field width and line segment S _idirection is vertical, claims Wei Xian region, this rectangular area R _i, if black region frame in Fig. 5 is line region;

2.a.2) by 8 intervals of 0 °～180 ° average deciles of direction, be respectively [0 ° 11.25 °) [168.75 ° of ∪, 180 °), [11.25 °, 33.75 °), [33.75 °, 56.25 °), [56.25 °, 78.75 °), [78.75 °, 101.25 °), [101.25 °, 123.75 °), [123.75 °, 146.25 °), [146.25 °, 168.75 °), and these 8 interval direction point abundances are turned to 0 ° 22.5 ° 45 ° 67.5 ° 90 ° 112.5 ° 135 ° 157.5 °, according to this quantization method, to all sketch line segment S _idirection quantize, establish sketch line segment S _iside vector turn to

2.a.3) according to 2b) quantized directions of sketch line segment Si in step get four direction (direction is mould 180 all), calculates sketch line segment S _icorresponding line region R _ifour direction gray level co-occurrence matrixes

2.a.4) according to 2.c) sketch map line segment S in step _igray level co-occurrence matrixes , calculate sketch map line segment S _igray level co-occurrence matrixes second-order statistic: entropy, contrast and correlativity, its computing formula is as follows:

Entropy: ${\mathrm{ENT}}_h=-\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}{G}_h(p,q)\log G_h(p,q)$

Contrast: ${\mathrm{CON}}_h=\underset{n=0}{\overset{K-1}{\mathrm{\Σ}}}{n}^2\underset{|p-q|=n}{\mathrm{\Σ}}{G}_h(p,q)$

Correlativity:

${\mathrm{COR}}_h=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}\frac{(p\·q)G_h(p,q)-u_p{u}_q}{s_p{s}_q}$

$\begin{array}{cc}{u}_p=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}p{G}_h(p,q)& u_q=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}q{G}_h(p,q)\end{array}$

$\begin{array}{cc}{S}_p^2=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}{G}_h(p,q){(p-u_p)}^2& S_q^2=\underset{p=1}{\overset{K}{\mathrm{\Σ}}}\underset{q=1}{\overset{K}{\mathrm{\Σ}}}{G}_h(p,q){(q-u_q)}^2\end{array}$

Wherein K represents the size of gray level co-occurrence matrixes, for sketch line segment S _idirection quantification amount.And according to the order of this four direction is spliced into the proper vector of one 12 dimension;

3.b) utilizing line segment co-occurrence matrix feature to use K-means clustering method is two classes by line segment cluster.

3.c) calculate sketch line segment S _ik-nearest neighbor distance a _i, computing formula is as follows:

$\begin{array}{cc}{a}_i=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{D}_{\mathrm{ij}}& D_i=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\end{array}$

Wherein (x _i, y _i), (x _j, y _j) be respectively sketch line segment S _i, S _jmiddle point coordinate, M represents and line segment S _ithe number of phase neighbour's line segment, Di _jrepresent line segment S _i, S _jeuclidean distance;

3.d) the average K-nearest neighbor distance Ma of calculating same class graticule section _i, i=1,2, computing formula is as follows:

${\mathrm{Ma}}_i=\frac{1}{n_i}\underset{j=1}{\overset{n_i}{\mathrm{\Σ}}}{a}_j,i=\mathrm{1,2}$

Wherein n _ithe number of same class graticule section.

3.e) sketch line segment little average K-nearest neighbor distance is designated as to W2 class line segment, portrays the line segment that light and shade grey scale change produces, and the average large sketch line segment of K-nearest neighbor distance is designated as W2 class line segment, portrays border and line target.

Described step (3), realizes sketch map Spatial Semantics according to the cluster analysis of sketch line segment classification results and sketch line segment spatial neighbors and divides, the areal map that the extraction in sketch map is more abstract.To the line segment in W2, analyze the aggregation of sketch line segment spatial neighbors, according to the statistical distribution of line segment aggregation extent, in sketch map, extract the areal map of more abstract aggregation zone, to the line segment in W1, analyze sketch line segment positional information, in sketch map, extract the areal map of non-aggregation zone; Specific implementation step is as follows:

3a) calculate the line segment in W2 aggregation W2a _i, computing formula is as follows:

$\begin{array}{cc}{W2a}_i=\frac{1}{n_i}\underset{j=1}{\overset{K}{\mathrm{\Σ}}}{D}_{\mathrm{ij}}& D_{\mathrm{ij}}=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}\end{array}$

Wherein (x _i, y _i), (x _j, y _j) be respectively W2 middle conductor middle point coordinate, k represents and line segment S _ithe number of phase neighbour's line segment, Di _jrepresent line segment S _i, S _jeuclidean distance;

3b) to the line segment in W2 aggregation W2a _icarry out statistics with histogram, according to the line segment in histogrammic peak-peak point W2 the interval R of optimum aggregation numerical value:

R＝[P-δ,P+δ]＝[L,U]，

Wherein, P is that in aggregation numerical value histogram, peak-peak is put corresponding aggregation numerical value, and δ is systematic parameter, and value is lower bound, the upper bound that 4, L, U represent respectively the interval R of optimum aggregation numerical value;

3c) choose line segment that in W2, aggregation equals optimum aggregation numerical value P as seed line-segment sets { E _k, k=1,2 ..., m}; If line segment E _kbe not added into certain line segment aggregate, with line segment E _kfor basic point recursively solves new line segment aggregate { ES _i, i=1,2 ..., l}, l <=m;

3d) actionradius is the circular primitive of the interval upper bound U of optimum aggregation numerical value, first the line segment in line segment aggregate is expanded, and then corrodes outward at line segment aggregate, extracts the areal map of aggregation zone in sketch map;

4.b) use a circular configuration element that radius is U, W1 line segment is expanded, be extracted in the region of the non-aggregation zone of sketch map.

Described step (4), becomes aggregation zone according to areal map by former SAR image mapped, determines non-aggregation zone and wireless segment region according to the value of pixel in the areal map of aggregation zone; Specific implementation step is as follows:

4.a) according to the value of pixel in the areal map of aggregation zone, if in areal map the value of pixel be 0, in corresponding SAR image, this pixel does not belong to aggregation zone, otherwise in corresponding SAR image, this pixel belongs to aggregation zone.The value of all pixels of judging area figure, extracts the aggregation zone in SAR image.

4.b) according to the value of pixel in the areal map of non-aggregation zone, if in areal map the value of pixel be 0, in corresponding SAR image, this pixel does not belong to non-aggregation zone, otherwise in corresponding SAR image, this pixel belongs to non-aggregation zone.The value of all pixels of judging area figure, extracts the non-aggregation zone in SAR image.。

4.c) remaining area of removing aggregation zone and non-aggregation zone in SAR image is called to wireless segment region.

Described step (5), is used watershed algorithm to carry out prime area division to SAR image, and SAR image is divided into many super pixels; Specific implementation step is as follows:

5a) the ratio gradient of calculating SAR image, computing method are as follows:

Centered by each pixel on image, use the window that size is 7 × 7, the ratio of grey scale pixel value in 0 °, 90 °, 45 ° and 135 °, direction in calculation window, wherein maximum ratio is exactly the final ratio response value of window center pixel, and gradient response is quantified as to [0,255] numerical value between, obtains final gradient map;

5b), according to carrying out watershed segmentation in the gradient map of SAR image, SAR image is divided into many super pixels; " Topographic distance and watershed lines " paper that dividing ridge method is delivered referring to Fernand Meyer.

Described step (6), respectively to aggregation zone, the super pixel in non-aggregation zone and wireless segment region adopts different consolidation strategies to merge, then integrates trizonal amalgamation result, obtains the final region division result of SAR image; Specific implementation step is as follows:

5a) for aggregation zone, the super pixel neighbouring relations of super pixel direct basis that belong in aggregation zone are merged, obtain aggregation zone division result { ER _i, i=1 ... p}, wherein p is the piece number that region is divided.

5b) for non-aggregation zone, will belong to the super pixel of non-aggregation zone, merge according to the constraint of gray scale difference between super pixel and w1 line segment direction, obtain the segmentation result { TR of non-aggregation zone _i, i=1 ..., q}, wherein q is the piece number that region is divided.

5c) for wireless segment region, will belong to the super pixel of wireless segment, will surpass pixel merging according to average gray and the variance of super pixel, obtain the segmentation result { NR in wireless segment region _i, i=1 ..., μ }, wherein μ is the piece number that region is divided.

5d) by non-step aggregation zone TR _icorresponding adjacent wireless segment region NR _ithe poor G of average gray in region, if G>=T, by region TR _ibe fused to and region TR _iadjacent aggregation zone ER _iin; Otherwise by region TR _ifusion can not sketch region NR _iin.

Finally will remain untreated non-aggregation zone as line target, extract again the sketch line segment that sketch line segment periphery correspondence in non-gathering line segment does not exist border, it is line segment length that position on corresponding sketch line segment SAR image is got long, and wide is that the region of 3 pixels is as line target.

Described step (7), to calculating the average gray value mG of each final zoning _i, use AP cluster to determine region class mark to zoning, finally obtain SAR image segmentation result, its AP clustering parameter is set: by two region R _iand R _jthe Euclidean distance of average gray value as the similarity si in two regions _j, region R _ireference degree get region R _iwith other Regional Similarities si _j, j=1 ..., the intermediate value of h.Obtain final SAR image segmentation result.

Advantage of the present invention is further illustrated by data and the image of following emulation.

1. simulated conditions

(1) in emulation experiment, choose Ku wave band 1m resolution PIPERIVER image, as shown in Figure 3.

(2) in the sketch map of the former SAR image obtaining in emulation experiment, have 1362 line segments, as shown in Figure 4.

(3) the aggregation K calculating in emulation experiment gets 5.

(4) the width size 2*HS+1 that extracts line region in emulation experiment is 2*20+1.

2. emulation content and result

Emulation 1, according to the local field of the line segment gray level co-occurrence matrixes feature of the present invention's definition, to Fig. 6 (a)-(j) middle sketch line segment (being white line segment in figure) extracts line segment feature, to verify the validity of feature.Sketch line segment in wherein Fig. 6 (a)-(e) is the aggregation zone line segment that belongs to PIPERIVER image, and give respectively its figure grade 1-5, sketch line segment in Fig. 6 (f)-(j) is the non-aggregation zone line segment cluster that belongs to PIPERIVER image, and gives respectively its figure grade 8-12.As shown in Figure 7, wherein X-axis is that figure grade, Y-axis are that intrinsic dimensionality, Z axis are eigenwert to the feature of sketch line segments extraction.Fig. 8 is sketch line segment classification results figure, wherein ater line segment is to portray the line segment that light and shade grey scale change produces, light gray line segment is the line segment of portraying border and line target, show that by experiment the feature that the present invention defines is effective, can effectively sketch line segment be divided into two class W1 and W2, W1 is the line segment of portraying border and line target, and W2 portrays the line segment that light and shade grey scale change produces.

Emulation 2, is divided into two class W1 and W2 according to line segment gray level co-occurrence matrixes feature line segment, and W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces; To the line segment in W2, analyze the aggregation of sketch line segment spatial neighbors, according to the statistical distribution of line segment aggregation extent, in sketch map, extract more abstract areal map; Result is as shown in Fig. 9 (a), and its white portion is the aggregation zone in sketch map.Expand according to W1 line segment, be extracted in the areal map of the non-aggregation zone of sketch map, result is as Fig. 9 (b).According to areal map, former SAR image mapped is become to aggregation zone, non-aggregation zone and wireless segment region, its result is respectively Figure 10 (a), Figure 10 (b), Figure 10 (c).

Emulation 3, obtains on the areal map basis in sketch map, this areal map being become to aggregation zone by former SAR image mapped, non-aggregation zone and wireless segment region in emulation 2.Use dividing ridge method that SAR image is carried out to over-segmentation, segmentation result is as Figure 11, respectively to aggregation zone, super pixel in non-aggregation zone and wireless segment region adopts different consolidation strategies to merge, as Figure 12, wherein Figure 12 (a), Figure 12 (b), Figure 12 (c) are respectively aggregation zones, the super pixel amalgamation result in non-aggregation zone and wireless segment region.Obtain the final area division result of SAR image, result as shown in figure 13, wherein Figure 13 (a) is the zoning plan of SAR image in patent " based on the SAR image partition method of semantic information classification ", Figure 13 (b) is respectively the zoning plan that the present invention extracts SAR image, and its white wire is zone boundary.

From Figure 13, the SAR image-region division figure that utilizes the present invention to extract; Comparing the details that obtains SAR image-region division figure in patent " based on the SAR image partition method of semantic information classification ", to describe (as line target) more accurate, and the consistent connectedness of ground object area marking off is higher, illustrate that ground object target that the present invention can be all on SAR image assembles the ground object area that forms and can extract preferably as forest.

Emulation 4, utilizes the PIPERIVER image of Ku wave band 1m resolution, uses the present invention to two width SAR Image Segmentation Usings.

The object of this experiment is can effectively to have solved because the special image-forming mechanism of SAR causes atural object aggregation zone and present regular light and shade grey scale change at SAR image S the present invention in order to verify, and causes similar region to be divided into the problem of multiclass.As shown in figure 14, its segmentation area class middle different gray scale of marking on a map represents inhomogeneity mark to experimental result.

Show that by emulation experiment the present invention can be partitioned into the atural object aggregation zone with light and shade grey scale change of consistent connectedness, effectively solve because the special image-forming mechanism of SAR causes atural object aggregation zone and presented regular light and shade grey scale change at SAR image, and caused similar region to be divided into the problem of multiclass.And can obtain on the whole the satisfied segmentation result of SAR image.

In sum, the present invention is a kind of SAR image partition method based on line segment gray scale symbiosis feature and region, can effectively solve because the special image-forming mechanism of SAR causes atural object aggregation zone and present regular light and shade grey scale change at SAR image, and cause similar region to be divided into the problem of multiclass.And can obtain on the whole the satisfied segmentation result of SAR image.

Claims (5)

1. the SAR image partition method based on line segment co-occurrence matrix feature and areal map, concrete steps comprise as follows:

(1) according to the sketch map of initial sketch model extraction SAR image;

(2) according to line segment gray level co-occurrence matrixes feature, line segment is divided into two class W1 and W2, W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces;

(3) realize sketch map Spatial Semantics according to the cluster analysis of sketch line segment classification results and sketch line segment spatial neighbors and divide, the areal map that the extraction in sketch map is more abstract;

(4) according to areal map, former SAR image mapped is become to aggregation zone, non-aggregation zone and wireless segment region;

(5) use watershed algorithm to carry out prime area division to SAR image, SAR image is divided into many super pixels;

(6) respectively to aggregation zone, the super pixel in non-aggregation zone and wireless segment region adopts different consolidation strategies to merge, then integrates trizonal amalgamation result, obtains the final region division result of SAR image;

(7) according to the average gray of the final zoning of SAR image, use AP cluster to determine region class mark to zoning, finally obtain SAR image segmentation result.

2. SAR image partition method according to claim 1, what wherein step (2) was described is divided into two class W1 and W2 according to line segment gray level co-occurrence matrixes feature line segment, W1 portrays border and line target, W2 portray light and shade grey scale change produce line segment in line segment gray level co-occurrence matrixes feature extract as follows:

(2.a) by sketch map line segment S _ithe locus at place is mapped to SAR image, then with sketch map line segment S _icentered by correspondence position, getting a size at SAR image is L _i× rectangular area (2*HS+1), wherein the length of rectangular area is line segment S _ilength L _i, direction and line segment S that rectangular area is long _idirection is parallel, and rectangle region field width is fixed value 2*HS+1, the direction of rectangle region field width and line segment S _idirection is vertical, claims Wei Xian region, this rectangular area R _i;

(2.b) by 8 intervals of 0 ° 180 ° average deciles of direction, be respectively [0 ° 11.25 °) [168.75 ° of ∪, 180 °), [11.25 °, 33.75 °), [33.75 °, 56.25 °), [56.25 °, 78.75 °), [78.75 °, 101.25 °), [101.25 °, 123.75 °), [123.75 °, 146.25 °), [146.25 °, 168.75 °), and these 8 interval direction point abundances are turned to 0 ° 22.5 ° 45 ° 67.5 ° 90 ° 112.5 ° 135 ° 157.5 °, according to this quantization method, to all sketch line segment S _idirection quantize, establish sketch line segment S _iside vector turn to

(2.c) according to 2b) sketch line segment S in step _iquantized directions get four direction (direction is mould 180 all), calculates sketch line segment S _icorresponding line region R _ifour direction gray level co-occurrence matrixes

(2.d) according to 2.c) sketch map line segment S in step _igray level co-occurrence matrixes calculate sketch map line segment S _igray level co-occurrence matrixes second-order statistic: entropy, contrast and correlativity, its computing formula is as follows:

Entropy: ${\mathrm{ENT}}_h=-\underset{p=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{K}{\mathrm{\&Sigma;}}}{G}_h(p,q)\log G_h(p,q)$

Contrast: ${\mathrm{CON}}_h=\underset{n=0}{\overset{K-1}{\mathrm{\&Sigma;}}}{n}^2\underset{|p-q|=n}{\mathrm{\&Sigma;}}{G}_h(p,q)$

Correlativity:

${\mathrm{COR}}_h=\underset{p=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{K}{\mathrm{\&Sigma;}}}\frac{(p\&CenterDot;q)G_h(p,q)-u_p{u}_q}{s_p{s}_q}$

$u_p=\underset{p=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{K}{\mathrm{\&Sigma;}}}p{G}_h(p,q)$ $u_q=\underset{p=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{K}{\mathrm{\&Sigma;}}}q{G}_h(p,q)$

$s_p^2=\underset{p=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{K}{\mathrm{\&Sigma;}}}{G}_h(p,q){(p-u_p)}^2$ $s_q^2=\underset{p=1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{q=1}{\overset{K}{\mathrm{\&Sigma;}}}{G}_h(p,q){(q-u_q)}^2$

Wherein K represents the size of gray level co-occurrence matrixes, for sketch line segment S _idirection quantification amount.And according to the order of this four direction is spliced into the proper vector of one 12 dimension.

3. SAR image partition method according to claim 1, what wherein step (2) was described is divided into two class W1 and W2 according to line segment gray level co-occurrence matrixes feature line segment, and W1 portrays border and line target, and W2 portrays the line segment that light and shade grey scale change produces; Carry out as follows:

3.a) according to line segment gray level co-occurrence matrixes feature calculation method, calculate line segment gray level co-occurrence matrixes feature, utilizing this feature to use K-means clustering method is two classes by line segment cluster;

3.b) calculate sketch line segment S _ik-nearest neighbor distance a _i, computing formula is as follows:

$a_i=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{D}_{\mathrm{ij}}$ $D_i=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}$

Wherein (x _i, y _i), (x _j, y _j) be respectively sketch line segment S _i, S _jmiddle point coordinate, M represents and line segment S _ithe number of phase neighbour's line segment, D _ijrepresent line segment S _i, S _jeuclidean distance;

3.c) the average K-nearest neighbor distance Ma of calculating same class graticule section _i, i=1,2, computing formula is as follows:

${\mathrm{Ma}}_i=\frac{1}{n_i}\underset{j=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{a}_j,i=\mathrm{1,2}$

Wherein n _ithe number of same class graticule section.

3.d) sketch line segment little average K-nearest neighbor distance is designated as to W2 class line segment, portrays the line segment that light and shade grey scale change produces, and the average large sketch line segment of K-nearest neighbor distance is designated as W2 class line segment, portrays border and line target.

4. SAR image partition method according to claim 1, what wherein step (4) was described becomes aggregation zone according to areal map by former SAR image mapped, non-aggregation zone and wireless segment region; Carry out as follows:

4.a) according to the value of pixel in the areal map of aggregation zone, if in areal map the value of pixel be 0, in corresponding SAR image, this pixel does not belong to aggregation zone, otherwise in corresponding SAR image, this pixel belongs to aggregation zone.The value of all pixels of judging area figure, extracts the aggregation zone in SAR image.

4.b) according to the value of pixel in the areal map of non-aggregation zone, if in areal map the value of pixel be 0, in corresponding SAR image, this pixel does not belong to non-aggregation zone, otherwise in corresponding SAR image, this pixel belongs to non-aggregation zone.The value of all pixels of judging area figure, extracts the non-aggregation zone in SAR image.。

4.c) remaining area of removing aggregation zone and non-aggregation zone in SAR image is called to wireless segment region.

5. SAR image partition method according to claim 1, wherein the described use AP cluster of step (7) is determined the class mark of zoning, finally obtains the segmentation result of SAR image, is the average gray value mG that calculates each final zoning _i, use AP cluster to determine region class mark to zoning, finally obtain SAR image segmentation result, its AP clustering parameter is set: by two region R _iand R _jthe Euclidean distance of average gray value as the similarity s in two regions _ij, region R _ireference degree get region R _iwith other Regional Similarities s _ij, j=1 ..., the intermediate value of h.Obtain final SAR image segmentation result.