CN104794730A - Superpixel-based SAR image segmentation method - Google Patents

Abstract

The invention discloses a superpixel-based SAR image segmentation method, mainly for solving the problems in the prior art that the computation complexity is high and tiny targets cannot be distinguished. The method comprises the following steps: 1, inputting an SAR image, namely completing the input of a to-be-segmented SAR image and acquiring image information; 2, generating superpixels for the input SAR image to obtain a superpixel image; 3, extracting the texture characteristics and spatial characteristics of the superpixel image; and 4, clustering the texture characteristics, combining the superpixels according to the spatial characteristics, and outputting the final segmentation result of the SAR image. The method disclosed by the invention has the advantages of effectively reducing the computation complexity of the traditional algorithm, shortening the SAR image segmentation processing time, distinguishing tiny targets and improving the segmentation accuracy, and can be applied to processing of images of airfield runways, farmland distribution and geological prospecting.

Description

Based on the SAR image segmentation method of super-pixel

Technical field

The invention belongs to technical field of image processing, particularly relate to image partition method, may be used for airfield runway, the image procossing of farmland distribution and geologic prospecting.

Background technology

Synthetic-aperture radar SAR is a kind of Coherent Imaging RADAR being operated in microwave region.It becomes the important means of current remote sensing observations with its high resolving power and round-the-clock, round-the-clock, large-area data retrieval capabilities, be widely used in resource, environment, archaeology and military affairs etc.SAR image comprises abundant target classification, but only may be interested in subregion wherein when image understanding.Area-of-interest is different along with the difference of application purpose.Such as, when detecting flood, interested region is waters, and in military activity, bridge near military base, road all may become interested target, are therefore partitioned into these target regions, effectively can not only reduce the calculated amount of computing machine, improve the real-time of algorithm, and identify that target is significant for correct.SAR image segmentation is that SAR image is understood and the key problem of decipher and difficult point place always.

SAR image comes from electromagnetic back scattering, therefore SAR image exists a large amount of coherent speckle noises, and this makes each pixel and its actual value often differ greatly, and when therefore conventional partitioning algorithm is applied to SAR image, general effect is not very desirable.Can say, coherent speckle noise is the key factor affecting SAR image segmentation quality.At present, main existence three kinds of methods reduce the impact of coherent speckle noise on Iamge Segmentation: the statistical model 1, setting up coherent speckle noise; 2, multiple dimensioned SAR image segmentation method is carried out; 3, SAR image is carried out to the pre-service of noise reduction.

When splitting natural image, often adopt the probability model of Gauss model as image of additivity, and for SAR image, due to the difference of imaging mechanism and the existence of coherent speckle noise, the additive Gaussian model in natural image can not be used to represent the statistical distribution characteristic of SAR image.Therefore, some Statistical Probabilistic Models of natural image segmentation are widely used in, as Markov random field model, Bayesian model, when splitting SAR image, hypothetical probabilities distribution meets Rayleigh distribution, Gamma distributes, K distribution etc., see Song Jianshe, Zheng Yongan, and Yuan Lihai, " diameter radar image is understood and application ", Beijing: Science Press, 2008.These algorithms all achieve reasonable segmentation effect, but these class methods operate single pixel due to it, so computation complexity is often comparatively large, still not can solve the impact of coherent speckle noise on segmentation result.

The representative of second method is as the SAR image partitioning algorithm of multiscale analysis model, the core concept of such algorithm SAR image is regarded as the texture of different scale, this character of different texture is presented according to different classes of target, by the segmentation to SAR image, be converted into the identification to different texture in SAR image.The algorithm quantity applying this method is a lot, as U.Kandaswamy etc. uses co-occurrence matrix to split SAR image, see U.Kandaswamy, D.A.Adjeroh, and M.C.Lee, " Efficient texture analysis of SAR imagery ", IEEE Transactions onGeoscience and Remote Sensing, vol.43, no.9, pp.2075-2083, 2005.X.Zhang uses the energy feature of co-occurrence matrix and wavelet decomposition to split SAR image, see X.Zhang, L.Jiao, F.Liu et al., " Spectral clustering ensemble applied to SAR image segmentation ", IEEE Transactions onGeoscience and Remote Sensing, vol.46, no.7, pp.2126-2136, 2008. Hou Biao etc. use second generation Bandelet territory Hidden markov tree model to split SAR image, see Hou Biao, Xu Jing, Liu Feng etc., " Iamge Segmentation based on second generation Bandelet territory Hidden markov tree model ", robotization journal, vol.35, no.5, 2009.But up to the present, still do not have a kind of method can to carry out effective modeling to whole texture, and such algorithm remain and process for single pixel, for the SAR image containing a large amount of pixel, the counting yield of this kind of algorithm is lower, and computing velocity is slow.

The third method is relatively more direct, namely before splitting SAR image, first it is carried out to the pre-service of noise reduction, to reduce the impact of coherent speckle noise on partitioning algorithm.In such, the simplest method is averaged by multiple pixel, and this is a very effective preprocess method, but its shortcoming is the marginal information that well can not keep image noise reduction while.

Summary of the invention

The object of the invention is to the deficiency overcoming above-mentioned prior art, propose a kind of SAR image segmentation method based on super-pixel, to effectively reduce computation complexity, shorten the time of segmentation, while noise reduction, well keep the marginal information of image, improve segmentation accuracy.

For achieving the above object, the invention provides the SAR image segmentation method based on super-pixel, comprise the steps:

1., based on a SAR image segmentation method for super-pixel, comprise the steps:

(1) to the SAR image I of input, its length L, width W and resolution R is extracted _i, and the interval R of resolution of minimum target is gone out according to Image estimations all in SAR image storehouse _s;

(2) super-pixel number N is estimated according to image parameter _s, generate the super-pixel S set={ s of SAR image to be split _i, wherein s _ibe i-th super-pixel i=1,2 ..., N _s;

(3) the textural characteristics F of super-pixel S set is calculated _g(i) and space characteristics F _n(i, j), wherein i, j=1,2 ..., N _sand i ≠ j;

(4) according to SAR image segmentation number K, use K-means algorithm to the textural characteristics F of super-pixel S set _gi () carries out cluster;

(5) according to the textural characteristics F of super-pixel S set _gi the cluster result of (), to two super-pixel s any in super-pixel S set _iand s _j, use Nearest neighbor rule, judge the whether corresponding same cluster of super-pixel, if s _iand s _jcorresponding same cluster, and their space characteristics F _n(i, j)=1, then merged, and the textural characteristics F to super-pixel S set _g(i) and space characteristics F _n(i, j) upgrades; Otherwise, do not merge;

(6) the super-pixel number N of the super-pixel S set after statistical updating _s(t), and by this super-pixel number N _s(t) and last statistics N _s(t-1) compare: if N _s(t)==N _s(t-1), then meet stop condition, export SAR image segmentation result; If N _s(t) < N _s(t-1), then do not meet stop condition, return step (4).

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

1. the present invention takes into full account that SAR image contains a large amount of coherent speckle noises, using the elementary cell of the super-pixel of the yardstick larger than single pixel as Iamge Segmentation, effectively reduces the impact of SAR image coherent speckle noise on segmentation result;

2. the present invention fully takes into account SAR image enormous size, containing a large amount of pixel, segmentation work is made to face a large amount of problem calculated, using the elementary cell of super-pixel as segmentation, and the textural characteristics extracted and space characteristics have simple to operate, the advantage that operand is little, effectively reduces the computation complexity of partitioning algorithm, shortens the processing time of Iamge Segmentation;

3. the present invention uses adaptive super-pixel merging method, well keeps the marginal information of image while noise reduction, improves segmentation accuracy.

Accompanying drawing explanation

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

Fig. 2 super-pixel image that to be the present invention produce containing the SAR image of organic field a width;

Fig. 3 is the super-pixel image that the present invention produces the SAR image that a width contains river and differ ent vegetation;

Fig. 4 is the super-pixel image that the present invention produces the SAR image that a width contains organic field and buildings;

Fig. 5 is the original SAR image containing organic field, buildings and vegetation that emulation experiment of the present invention adopts;

Fig. 6 is the result figure split the emulation of Fig. 5 with the present invention;

Fig. 7 is the original SAR image containing river, the woods and farmland that emulation experiment of the present invention adopts;

Fig. 8 is the result figure split the emulation of Fig. 7 with the present invention.

Embodiment

With reference to Fig. 1, the present invention is based on the SAR image segmentation method of super-pixel, comprise the steps:

Step 1: according to the SAR image of input, obtains image information.

Image I all in input SAR image storehouse, and the SAR image I to input, extract the length L of every piece image, width W and resolution R _i, the interval R of resolution of minimum target is gone out according to these parameter estimation _s.

Step 2: estimate super-pixel number, obtains super-pixel set.

2a) according to length L, the width W of piece image, resolution R _ir interval with the resolution of minimum target _s, the super-pixel number N of this width image is estimated by the following two kinds situation _s:

The first situation is: according to the input length L of SAR image to be split, width W and resolution R _i, estimate super-pixel number N _sinterval:

$N_s\&Element;\frac{L\&times;W\&times;R_I^2}{R_s^2}\&cap;\frac{L\&times;W}{\mathrm{Num}\left(s\right)},$

Wherein, the interval of the number of pixels of Num (s) contained by each super-pixel, interval is [50,200], super-pixel number N _sfor appointing a round values of getting in interval;

The second situation is: as the interval R of the resolution being unable to estimate out minimum target _stime, according to length L, the width W of input SAR image to be split, estimate super-pixel number N _sinterval be:

$N_s\&Element;\frac{L\&times;W}{\mathrm{Num}\left(s\right)};$

2b) generate the super-pixel S set={ s of SAR image to be split _i, wherein s _ibe i-th super-pixel i=1,2 ..., N _s:

2b1) obtain the super-pixel number N of SAR image _safter, inner in SAR image to be split, uniform design N _sindividual pixel seed, and centered by these seeds, use the method for level set to allow the level set movements equation that these seeds expand according to following seed expand:

Φ ⁿ⁺¹＝Φ ⁿ+F|▽Φ ⁿ|Δt

Wherein Φ ⁿrepresent the level set function after developing for n-th time, | ▽ Φ ⁿ| represent the Grad of the level set function after developing for n-th time, F represents velocity function, and Δ t represents the time interval of curve evolvement;

2b2) when seed-bearing edge all overlaps, just stop the expansion of seed, each seed is a super-pixel s _i, thus complete input SAR image super-pixel S set={ s _igeneration work.

Step 3: the textural characteristics F calculating super-pixel S set _g(i) and space characteristics F _n(i, j).

3a) for any one super-pixel s _i, by the textural characteristics F of following formulae discovery super-pixel S set _g(i):

$F_g\left(i\right)=\frac{H(i,1:m)}{\mathrm{Num}\left(s_i\right)},$

Wherein, H (i, 1:m) is super-pixel s _igrey level histogram, m represents the number of greyscale levels of image, Num (s _i) represent super-pixel s _icontained number of pixels;

3b) for any two super-pixel s _iand s _j, by the space characteristics F of following formulae discovery super-pixel S set _n(i, j):

Step 4: cluster is carried out to textural characteristics, and in conjunction with space characteristics, super-pixel is merged.

The existing clustering method to textural characteristics has multiple, such as K-means, fuzzy C-mean algorithm method, Nearest neighbor rule etc., and this example adopts but is not limited to K-means.

Being implemented as follows of this step:

4a) for any two super-pixel s _iand s _j, calculate it respectively to all cluster centre c according to variable volume distance VBSD _ndistance, dist (s _i, c _n), dist (s _j, c _n), wherein n=1,2 ..., K;

4b) by super-pixel s _ithe cluster centre corresponding apart from minimum value distributes to i-th super-pixel s _i, as super-pixel s _icorresponding cluster c ⁱ:

$c^i=\underset{n}{\arg }\min \left(\mathrm{dist}(s_i,c_n)\right);$

4c) by super-pixel s _jthe cluster centre corresponding apart from minimum value distributes to a jth super-pixel s _j, as super-pixel s _jcorresponding cluster c ^j:

$c^j=\underset{n}{\arg }\min \left(\mathrm{dist}(s_j,c_n)\right)$

4d) judge this two super-pixel s _iand s _jwhether corresponding identical cluster centre, if these two super-pixel s _iand s _jcorresponding cluster centre is identical, then super-pixel s _iand s _jbelong to same cluster, otherwise, do not belong to same cluster;

4e) judge the super-pixel s of same cluster _iand s _jspace characteristics, if their space characteristics F _n(i, j)=1, then by super-pixel s _iand s _jmerge, obtain new super-pixel S set ^*otherwise, do not merge.

Step 5: to new super-pixel S set ^*space characteristics and textural characteristics upgrade.

5a) at two super-pixel s _iand s _jafter merging, the sequence number i of one of two super-pixel is removed, represent the super-pixel after merging with the sequence number j of another super-pixel, obtain new textural characteristics F _g' (j):

$F_g^{\&prime;}\left(j\right)=\frac{F_g\left(i\right)\&times;\mathrm{Num}\left(s_i\right)+F_g\left(j\right)\&times;\mathrm{Num}\left(s_j\right)}{\mathrm{Num}\left(s_i\right)+\mathrm{Num}\left(s_j\right)},$

Wherein F _gi () represents the front super-pixel s of merging _itextural characteristics, F _gj () represents the front super-pixel s of merging _jtextural characteristics, Num (s _i) represent the front super-pixel s of merging _ithe number of pixels comprised, Num (s _j) represent the front super-pixel s of merging _jthe number of pixels comprised;

5b) delete the front s of merging _itextural characteristics F _gi (), with the new textural characteristics F produced _g' (j) substitutes s _jtextural characteristics after merging.

Step 6: super-pixel amalgamation result is differentiated.

Statistics is each merges rear new super-pixel S set ^*super-pixel number N _s(t), and by this super-pixel number N _s(t) and last statistics N _s(t-1) compare: if N _s(t)==N _s(t-1), then meet stop condition, export SAR image segmentation result; If N _s(t) < N _s(t-1), then do not meet stop condition, return step (4), continue to merge super-pixel.

Effect of the present invention can be further illustrated by the following emulation experiment to true SAR image:

1, emulation experiment condition

Emulation of the present invention is at windows XP, and SP2, CPU Pentium (R) 4, basic frequency 2.4GHZ, software platform is that Matlab7.0.1 realizes.The true SAR image that emulation experiment is selected is the SAR image of two width Ku wave bands, as shown in figure 5 and figure 7, wherein Fig. 5 is the SAR image on California, United States airport, is of a size of 522 × 446, resolution is 3 meters, containing terrain object such as organic field, buildings and vegetation; Fig. 7 is the width SAR image near A Erbukeerke city, New Mexico, is of a size of 600 × 432, and resolution is 1 meter, containing terrain object such as river, the woods and farmlands.

2, content and result is emulated

Emulation one, the super-pixel produced in the SAR image of a width containing organic field with the present invention, forms super-pixel image as shown in Figure 2.

Emulation two, produces super-pixel in the SAR image containing river and differ ent vegetation at a width with the present invention, forms super-pixel image as shown in Figure 3.

Emulation three, contains the super-pixel produced in the SAR image of organic field and buildings to a width with the present invention, form super-pixel image as shown in Figure 4.

Emulation four, carry out emulation experiment segmentation with the present invention to Fig. 5, result is as Fig. 6.

As can be seen from Figure 6, the present invention can locate the edge of airfield runway accurately, and runway and other terrain object is separated accurately.For the buildings of near airports, the present invention can be found, comes even if the terrain object of itself and surrounding also can distinguish by indivedual isolated buildingss.For ground vegetation, the vegetation region with different visual characteristic can effectively separate by the present invention.

Emulation five, carry out emulation experiment segmentation with the present invention to Fig. 7, result is as Fig. 8.

As can be seen from Figure 8, the present invention accurately can locate the edge in river course and the edge of road, and is separated by the different vegetation on ground, and the present invention can also effectively tell special target, as the trees on meadow, the island etc. in river.For same class terrain object, the segmentation result of gained of the present invention has good region consistency, and wrong point seldom.

As can be seen from above the simulation experiment result, the present invention can not only produce super-pixel image, and can effectively be split SAR image by the mode based on super-pixel, the edge of Different Ground target accurately can be located when splitting, the region consistency of similar target can be kept, and can locate and find special objective isolated in SAR image accurately.

Claims (6)

1., based on a SAR image segmentation method for super-pixel, comprise the steps:

(1) to the SAR image I of input, its length L, width W and resolution R is extracted _i, and the interval R of resolution of minimum target is gone out according to Image estimations all in SAR image storehouse _s;

(2) super-pixel number N is estimated according to image parameter _s, generate the super-pixel S set={ s of SAR image to be split _i, wherein s _ibe i-th super-pixel i=1,2 ..., N _s;

(3) the textural characteristics F of super-pixel S set is calculated _g(i) and space characteristics F _n(i, j), wherein i, j=1,2 ..., N _sand i ≠ j;

(4) according to SAR image segmentation number K, use K-means algorithm to the textural characteristics F of super-pixel S set _gi () carries out cluster;

(5) according to the textural characteristics F of super-pixel S set _gi the cluster result of (), to two super-pixel s any in super-pixel S set _iand s _j, use Nearest neighbor rule, judge the whether corresponding same cluster of super-pixel, if s _iand s _jcorresponding same cluster, and their space characteristics F _n(i, j)=1, then merged, and the textural characteristics F to super-pixel S set _g(i) and space characteristics F _n(i, j) upgrades; Otherwise, do not merge;

(6) the super-pixel number N of the super-pixel S set after statistical updating _s(t), and by this super-pixel number N _s(t) and last statistics N _s(t-1) compare: if N _s(t)==N _s(t-1), then meet stop condition, export SAR image segmentation result; If N _s(t) < N _s(t-1), then do not meet stop condition, return step (4).

2. the SAR image segmentation method based on super-pixel according to claims 1, wherein said step (2) estimate super-pixel number N according to image parameter _s, undertaken by the following two kinds situation:

The first situation is: according to the input length L of SAR image to be split, width W and resolution R _i, estimate super-pixel number N _sinterval:

$N_s\&Element;\frac{L\&times;W\&times;R_I^2}{R_s^2}\&cap;\frac{L\&times;W}{\mathrm{Num}\left(s\right)}$

Wherein, the interval of the number of pixels of Num (s) contained by each super-pixel, interval is [50,200], super-pixel number N _sfor appointing a round values of getting in interval;

The second situation is: as the interval R of the resolution being unable to estimate out minimum target _stime, according to length L, the width W of input SAR image to be split, estimate super-pixel number N _sinterval be:

$N_s\&Element;\frac{L\&times;W}{\mathrm{Num}\left(s\right)}.$

3. the SAR image segmentation method based on super-pixel according to claims 1, calculates the textural characteristics F of super-pixel S set in wherein said step (3) _gi (), for any one super-pixel s _i, by following formulae discovery:

$F_g\left(i\right)=\frac{H(i,1:m)}{\mathrm{Num}\left(s_i\right)}$

Wherein, H (i, 1:m) is super-pixel s _igrey level histogram, m represents the number of greyscale levels of image, Num (s _i) represent s _icontained number of pixels.

4. the SAR image segmentation method based on super-pixel according to claims 1, calculates the space characteristics F of super-pixel S set in wherein said step (3) _n(i, j), for any two super-pixel s _iand s _j, by following formulae discovery:

5. the SAR image segmentation method based on super-pixel according to claims 1, uses Nearest neighbor rule in wherein said step (5), judges the whether corresponding same cluster of super-pixel, carry out as follows:

5a) for any two super-pixel s _iand s _j, calculate it respectively to all cluster centre c according to variable volume distance VBSD _ndistance dist (s _i, c _n), dist (s _j, c _n), wherein n=1,2 ..., K;

5b) by super-pixel s _ithe cluster centre corresponding apart from minimum value distributes to i-th super-pixel s _i, as super-pixel s _icorresponding cluster c ⁱ:

$c^i=\underset{n}{\arg }\min \left(\mathrm{dist}(s_i,c_n)\right)$

5c) by super-pixel s _jthe cluster centre corresponding apart from minimum value distributes to a jth super-pixel s _j, as super-pixel s _jcorresponding cluster c ^j:

$c^j=\underset{n}{\arg }\min \left(\mathrm{dist}(s_j,c_n)\right)$

Judge super-pixel s _iand s _jwhether corresponding identical cluster centre, if super-pixel s _iand s _jcorresponding cluster centre is identical, then super-pixel s _iand s _jbelong to same cluster.

6. the SAR image segmentation method based on super-pixel according to claims 1, the textural characteristics F to super-pixel S set in wherein said step (5) _g(i) and space characteristics F _n(i, j) upgrades, and carries out as follows:

5d) at two super-pixel s _iand s _jafter merging, the sequence number i of one of two super-pixel is removed, represent the super-pixel after merging with the sequence number j of another super-pixel, obtain new textural characteristics F ' _g(j):

$F_g^{\&prime;}\left(j\right)=\frac{F_g\left(i\right)\&times;\mathrm{Num}\left(s_i\right)+F_g\left(j\right)\&times;\mathrm{Num}\left(s_j\right)}{\mathrm{Num}\left(s_i\right)+\mathrm{Num}\left(s_j\right)},$

Wherein F _gi () represents the front super-pixel s of merging _itextural characteristics, F _gj () represents the front super-pixel s of merging _jtextural characteristics, Num (s _i) represent the front super-pixel s of merging _ithe number of pixels comprised, Num (s _j) represent the front super-pixel s of merging _jthe number of pixels comprised;

5e) delete the front s of merging _itextural characteristics F _gi (), with the new textural characteristics F ' produced _gj () substitutes s _jtextural characteristics after merging.