Ant Colony Fuzzy ClusteringAlgorithm Applied SAR Image ...

Post on 01-Oct-2021

3 views 0 download

Transcript of Ant Colony Fuzzy ClusteringAlgorithm Applied SAR Image ...

Ant Colony Fuzzy Clustering Algorithm Applied toSAR Image Segmentation

Li Chunmao', Wang Lingzhi2, Wu Shunjun31,3.National Lab. of Radar Signal Processing, Xidian University; 2.Xi'an Institute of Post and Telecommunications,

Xi'an, Shannxi, 710071, P.R. China

stucatg 163 .com, wlzmary@sina.com, sjwu@xidian.edu.cn

Abstract A method of dynamic fuzzy clustering analysis

based on ant colony algorithm for SAR image segmentation

is proposed. The method confirms dynamically the

clustering number and center by the stronger fuzzy

clustering ability of ant colony algorithm. Texture feature of

SAR image is calculated according to gray level

co-occurrence matrix (GLCM), and the proper feature

vector is selected through statistic analysis. The

measurement SAR image segmentation experiment indicates

that the algorithm can segment the target fast and exactly,

and is an effective SAR image segmentation method.

Key words ant colony algorithm, fuzzy clustering, SAR

image segmentation, gray level co-occurrence matrix

I. INTRODUCTION

The interpreted SAR image in military applicationmainly encircles how to detect and identify syntheticobject from the image containing complex background,typical synthetic object is such as tank, vehicle, fleet andso on. To the detection and identification of this synthetictarget, common dealing method of SAR system firstlydetect interesting target area fast in the image containingcomplex background, called by ROI[1][2]. Then exact

segment is taken to obtain ROI in order to obtain theobject area from obtained ROI. Whether the segmentationis good will directly influence the characteristicsextraction of target and the target classification andidentification.

Ant colony algorithm[3][4] is a kind of bionicevolution one, and is a random searching method which isof discrete, parallel, robust, positive-going feedback andfuzzy clustering ability. It is successfully applied to suchassemble optimization question as traveler question andworkshop task dispatch. J.Casillas proposed an automaticstudy method for fuzzy rules using ant colony algorithm.The discrete and parallel characteristic of ant colony

algorithm is very practical to discrete SAR image. Thepath selection method based on probability is of wideapplication prospect in fuzzy clustering question.

II. ANT COLONY ALGORITHM

Ant Colony Algorithm is also called Ant Algorithm,which is a bionic evolution algorithm proposed by ItalianScholar M.Dorigo who is enlightened by path selectionbehavior of ants in theirs searching food process in 1992.It is found by observing that ants always can find anoptimal path to food source in searching food process.After the optimal path is interdicted, Ants can steer clearof obstacle quickly and find the optimal path again. Thekind ability of ants is produced through the informationexchange and mutual cooperation between ant colonies.Every ant releases a kind of information hormone duringrandom sashaying process, and this hormone volatilizesconstantly with time lasting. If there is more ants selectthis path, the hormone on this path will be strengthened.While every ant is of the ability to perceive thisinformation hormone strength, they will select the pathwith stronger information hormone, leading to added ants

selecting this path. Thus a positive-going feedback isformed.

A. The step ofAnt Colony Algorithm

The ecumenic step of ant colony algorithm appliedto solve problem is as follows:

1) Problem analysis: making the problem to besolved abstract, and giving the problem space parametersvariables specific implication.

2) Initialization: giving every variable the initialvalue, ants all wait in the hole for starting out to searchfood.

3) Optimal process: ants make a dynamicselection in accordance with given path length and

0-7803-9582-4/06/$20.00 c2006 IEEE

information hormone strength, and release informationhormone during moving.

4) Cessation conditions: if given conditions aremeet, The algorithm will cease; otherwise step onto 3.

Step 3 is an adaptive process, and it embodies theessence of ant colony algorithm: selecting mechanismand updating mechanism.

B. Mathematic Describe ofAnt Colony Fuzzy ClusteringAlgorithm

to the following formula:

phij (t) = pphij (t) +Aph (4)

Where, p is the waning extent of information

amount with time going on, Aphii is the augmentation

of path information amount in this circulation.

N

Aphi = E AphAk=l

Give the initial SAR image X, and look every pixel

Xj (j = 1,2,.* , N) as an ant, every ant stands for the

feature vector of the pixel. Image segmentation is the

process that these ants with different feature vector

searching food source. The distance of any pixel Xi to

Xi is di,, using euclidean distance to calculate:

m

di Z,Pk(Xik Xjk) (1)

k=l

Where, m is the number of feature vector, Pk is

weight factor, which is set in terms of the influence

extent of every feature vector of pixel to clustering.

r is set as clustering radius, ph as information

amount, Then:

phi = ldi.<r

The probability of the path Xito Xj isp:

ph (t)< (t)

pa

pha(t)<7 (t)scS

otherwise

j ES

Here, 17i (t) is apocalyptic

function, a , is respectively the influence

the accumulated information and apocalyptic

guida

factor

guida

tunction to path selection.

S = Xs dsj < r, s = 1,2, * , N} is ambulatory path

set.

With the ants moving, the information amount on

every path is changing. Through one circulation, the

information amount on every path is adjusted according

Aphfk is the information amount remained by the k

ant in this circulation.

C. Set ofthe Guidance Function

Guidance function embodies the resemblance extent

between the pixel and clustering center, and is expressedthrough the following formula:

1

111=-r

m

Pk(Xik Xjk)k=l

(6)

Here, r is clustering radius. The bigger theclustering radius is, and the bigger Guidance functionvalue is, with which the probability of this clusteringcenter becomes bigger.

III. SAR IMAGE FEATURES EXTRACTION

(2) A. Gray Level Co-occurrence Matrix

Texture gray level co-occurrence matrix[5] is a

statistics method based on 2-Order assembly condition

probability density function of image. It is different fromgray level statistics analysis, image pixel isn't considered

(3) separately, but the frequency between pixels with some

relations is described. Generally crudity and direction isthe most predominant characteristic when texture is

Lnce distinguished. Two characteristics are corresponding to

r of the step size d and direction 0 in gray level

ince co-occurrence matrix. Gray level co-occurrence matrix isdefined as follows:

P(h, k) = [p(h, k d, )] (7)

Its meaning is that the satisfied position condition on

the image is that step size is d and direction is 0, and

gray level is the emerging times of pixel h and pixel k

pairs at the same time. The paper select Contrast, Energy,

(5)

r~~~~~~~~~_.-._ __. _ __1 __

Entropy, Inverse Difference Moment as feature vector,defined as follows:

Contrast: CONTR = E (h - k)2 M(h, k) (8)h k

Energy: ASM = M(h,k)2 (9)h k

Entropy: ENTRO = E M(h, k) log M(h, k) (10)h k

Inverse Difference Moment:

IDM=ZZ M(h,k)h k (11)

M(h, k) is the normalization processing to

gray level co-occurrence matrix P(h, k).

M(h k)= ZP(h ,k) (12)

h k

B. Features Extraction

Selecting an 11 * 11 pixel window with step size as 1and direction as 0 degree, making statistic analysis toevery feature in the experimental area selected onexperiment image. The calculation of every feature isneeded to select several experiment samples. Result issuch as table I.

Table I. Typical Texture Feature Values

ASM CONTR ENTRO IDM

Field 0.0729 1.3000 -2.9617 0.6290

Village 0.0175 5.2424 -4.5546 0.4741

City 0.0849 2.6499 -3.5037 0.6568

Forest 0.1132 0.6396 -2.5506 0.7486

Road 0.1092 1.4383 -2.7940 0.6719

Runway 0.3402 0.3174 -1.3415 0.8426

Lawn 0.0869 1.0921 -2.7958 0.6598

different food source Cj according to (1). If di is

zero, then the membership of the pixel to this kind is 1,

otherwise if d.. < r, calculating guidance function

according to (6) and information amount to every path

through (4).

4) Calculating the membership of pixel according to

(3). Judge if he membership is bigger than A, if yes,

calculate information augmentation Aphi, according to

formula and update information amount. Update the j

kind clustering center referring to the following formula.

J is the number of element in the kind Cj:

(13)k=l

Otherwise, record the ant into set SS, SS is the pixel

assembly unclassified.

5) Calculating the distance of every kind. When

the distance is smaller than threshold 8, combine two

kinds into one, updating the new clustering center.

6) If there is still the pixel to be classified, return to

step 2. Otherwise end.

B. Experimental Results andAnalysis

IV. EXPERIMENT

A. Algorithm Flow

Initial SAR image as 256 colors gray chart, shown

in figure I(a) and figure 11(a).

1) Initializing the parameters a, ,6, phi , r .

2) Calculating every feature vector of pixel Xi in

accordance with (8)-(15).

3) Calculating the distance d f pixel Xi

Figure I(a). SAR image of airport runway(572 X 790)

Figure I(b). Result of segmentation

Figure 11(a). SAR image of fields(348 X 298)

Fig. I (a) and Fig. 11 (a) are two SAR images cut off,and they are respectively airport runway and field,resolution is 3 meter.

Fig. I (b) and Fig. II (b) are respectively the

segmentation results using ant colony fuzzy clustering

algorithm when a =1,/3 =1 , r = 50 X = 0.9,£ = 0.5. It is seen from figures that the proposed image

segmentation algorithm the detected edge is more

continuous, edge detail structure embodies good, and the

part with low gray level value is also detected, So the

segmentation result are more accurate. The segmentation

result in Fig. 11(b) is a little worse than that in Fig. 1(b) as

a result of speckle effect. So there will be a good result if

speckle reduction is taken before image segmentation.

V. CONCLUSION

Method of dynamic fuzzy clustering analysis basedon ant colony algorithm for SAR image segmentation isproposed in the paper. Experiment results verify theeffectivity of the algorithm. Ant colony algorithm are ofthe discrete, parallel, and fuzzy clustering abilities, andthese features are more applied to the further SARinformation extraction such as texture classification, edgedetection and so on.

REFERENCE

[1] Bhanu B, Dudgeon D E, Zelnio E G. Introduction tothe special issue on automatic target detection andrecognition. IEEE Trans on Image Processing, 1997,16(1) :1 - 6.

[2] Ross T D. SAR ATR: So what's the problem? AnMSTAR perspective. SPIE ,1999, 3721 :662 - 672.

[3] Dorigo M, Maniezzo V, Colorni A. Ant system:optimization by a colony of cooperating agents.IEEE Trans On System, Man, and Cybernetics,1996,26(1):29-41.

[4] Marco Dorigo, Gianni Di Caro, Luca M

Gambardella. Ant Algorithms for DiscreteOptimization. Proceedings of the Congress on

Evolutionary Computation, 1999.[5] Robert M H, Shanmugam K. Textural Features for

Image Classification. IEEE Transaction on System,Figure 11(b). Result of segmentation Man and Cybernetics, 1973,3(6):610-621.