Sei sulla pagina 1di 5

A

Simplex-Genetic Hybrid Approach for the Classification of Image Textures


Li Pan School of Remote Sensing Information and Engineenng, Wuhan University, 129 Luoyu Road, Wuhan, Hubei 430079, P.R.China E-mail: li.pan@126.com

School of Electronic Information Wuhan University 129 Luoyu Road, Wuhan, Hubei 430079, P.R.China E-mail: zhenghong@21cn.com combines the GA with the simplex method, and compared our hybrid approach with traditional GA. This paper is organized as follows: The simplex method is overviewed in Section 2. Section 3 describes the hybrid algorithm integrating simplex and GA. Section 4 presents the experimental results and conclusions are given in Section 5.
II. THE SIMPLEX METHOD

Hong Zheng

Abstract-This paper proposes a hybrid approach to classify image textures by integrating genetic algorithms and the simplex method. The simplex method is a kind of local searching method that gets new and better simplex points by reflection, expansion and contraction operations. Since the method converges quicldy, this paper employs the local search characteristic of the simplex method to avoid the premature of genetic algorithms. Based on the integration of genetic algorithms and the simplex method, a hybrid algorithm is proposed to discriminate image textures. The classification experiments on five classes of aerial images are presented for the purpose of the performance comparison with genetic algorithms. The experimental results show that the proposed method is feasibility and its performance is better than that of genetic algorithms.

I. INTRODUCTION

Genetic algorithms (GA) have been demonstrated to be a promising search and optimization technique. GA has been successfully applied to system identification and a wide range of applications including image processing scheduling , routing and control . One of the main obstacles in applying GA to complex problems has often been the high computational cost due to their slow convergence rate. The convergence rate of a GA is typically slower than that of local search techniques, because it does not use much local information to determine the most promising search direction. A common strategy in the literature for dealing with the GA's slow convergence problem is to combine the GA with a complementary local search technique [1J[5]. The rationale of such a strategy is that such a hybrid approach can combine the merits of the GA with that of a local search technique. Because of the GA, a hybrid approach is less likely to be trapped in a local optimum than a pure local search technique is. Due to its local search, a hybrid approach often converges faster than the pure GA does. Generally speaking, a hybrid approach usually can explore a Fig. 1. An example of the simplex method better tradeoff between computational cost and the global optimality of the solution found. Fig. 1 illustrates how a simplex method is applied in a In this paper, toward the objective of classifying aerial minimum optimization problem involving two variables. As image textures, we developed a hybrid approach that shown in Fig. 1 point xj,x2 and x3 form an original simplex.

The simplex method is a local search technique that uses the evaluation of the current data set to determine the promising search direction. It has two modes: the constrained simplex method and non-constrained simplex method 61. The later method is used to solve the problem of the texture classification in this paper. In the non-constrained simplex method, a simplex is defined by a number of points equal to one more than the number of dimensions of the search space. Firstly, the maximum point and minimum point are determined from the simplex, then a new better point is found by reflection, expansion and constriction operations. Secondly, the maximum point or minimum point is replaced by the new point and a new simplex is generated. Above procedure is repeated until the optimal solution is found.

0-7803-9422-4/05/$20.00 C2005 IEEE

1192

Point xl is the best evaluation (minimum point ) and point X3 is the worst evaluation (maximum point ). Point * represents the centroid of xi and x2, i.e. (xI+ x2)/2. Reflecting point X3 across Wgenerates x4 , which together with points xl- and x2, forms a new simplex. x4 is defined as

I=

x= x

+ a (x -X)

()

Where a is called refection coefficient (a>0), normally a =1. After reflection operation, three cases are discussed as follows: 1.If f(x4)<f(x1), this means reflection direction is good for decreasing function value, so a new point further along the reflection direction is generated using the equation
X X

+ r (x4 -X)

(2)

(7) The "texture energy" feature is invariant to rotation and Where , is the contraction coefficient, xais the best scale changes of textures. More details are given in Ref.[7]. In our approach, all elements in a mask form a chromosome point between x3 and X4. If f(x6) < f(xa), X3 is replaced by x6 and points x1,x2 and x6 of GA. Since the mask is a symmetric array and the sum of form a new simplex. If f(x6)>f1(xa), the contraction all elements is zero, only ten elements in a mask are used to operation is implemented to generate two new contraction form a chromosome of GA. The detailed steps for the hybrid algorithm is given as follows: point x7 and x8 using following equations: Step J:Initialize a population which includes 30 (4) decimal-coded chromosomes. Each chromo- some consists X7 X2 + (Xl-X)12 of ten elements in a mask. Step 2: Calculates the fitness value of each chromosome (5) using following equation: XS= X3+ (X-X)/2 As a result, point x1,x7 and x8 form a new simplex. Fitness = min d2 / max d, I (8) For above cases, any of them will generate better simplex. In other words, there must be a point in the new where d, is the inner-class square error of the energy of simplex, whose function value is less than the minimum all same texture images , d2 is a inter-class square error of function value in the old simplex. According to above rules, the energy of different class texture images. the simplex updating is repeated until the stopping According to the fitness value of each chromosome, conditions are met. chromosomes are sorted in a descending order. The 3 best From above analysis, it can be seen that the simplex chromosomes are copied to next generation. method is a kind of local optimization method. It can be Step 3: Divides the rest 27 chromosomes into 9 groups employed to speed the convergence of GA and prevent GA and each group consists of 3 chromosomes. In addition, add from being trapped in a local optimum. the best chromosome into each group so that each group forms a simplex including 4 points (chromosome). For each group, a new point is calculated by using the simplex m. THE SIMPLEX-GENETIC HYBIRD ALGORITHM method and saved as a individual in next generation. As a In this section, we describe the approach integrating the result, a total of 9 new points are saved as an individual in
X6 = x+ 1(x-x)

Where at is called the expansion coefficient. 1) If f(x5)<f(x4), X3 is replaced by x5 and points xj,x2 and x5 form a new simplex. If R(x5)Pf(x4), X3 is replaced by x4 and points x1,x2 and x4 form a new simplex. 2)If f(x1) < f(x4) < f(x2), x3 is replaced by x4 and points x1,x2 and x4 form a new simplex. 3)If f(x4)>f(x2), the contraction operation is implemented to generate a new contraction point x6 using following equation:

simplex method and the GA. In the approach, each individual in a population of GA is regarded as a point in a simplex. The fitness function value of each individual is also regarded as the function value of each point in the sitplex. According to the operation rules of the simplex method, the best point in each population is selected and grouped with the rest individuals to implement genetic operation such as cross and mutation operations. As a result, a new population is generated. The new population undertakes simplex operations and genetic operations again, and produces its children . In such way, the simplex method and GA are combined to update population continuously until the termination condition is met. In this paper, the hybrid approach is used to generate the "tuned" mask for the classification of image textures. The "tuned" mask is a kind of 5 by 5 symmetric array A(ij), it is convolved with a image I(ij) and generates following convolved image F(ij): (6) F(i, j) = A(i, j) * I(i, j)
The local "texture energy" can be evaluated at each pixel location (ij) in the convolved image F(ij) over a large window size w, X wy (w,=wy=9) as the measure of texture features in the spatial domain. The "texture energy" is defined as follows:

(3)

E(i, j) = (

2F 2(i j))/(WXw,)

1193

Best

Worst L 1V-Ranked population


Fig.2 The reproduction procedure of proposed hybrid algorithm

INew population

met.

Step 4: For current population, genetic algorithms are implemented to produce 18 children. Theses children are also saved as individuals in the next generation so as to ensure that the next generation still includes 30 individuals. Step 5: Repeat step 2-5 until the termination condition is
Fig.2 shows the reproduction procedure of the proposed algorithm.
IV. EXPERIMNiTAL RESULTS

next generation.

number, fb, and fbmin represent the maximum and minilimum fitness value in a simplex respectively.

In this section we present experimental results of learning "tuned" mask using genetic algorithms and the simplex method (GASPX). A total of 50 texture images with 100 by 100 size are used to learn the "tuned" mask. These images consist of residential lands, bushes, cultivated lands, mountain lands, and dry lands, which are selected from real aerial images. Some of them are shown in Fig. 3. The learned mask is tested on 47 bush images, 23 resident land images, 13 cultivated land images, 20 mountain land images and 15 dry land images. In our experiments, a new point in a simplex is generated by following steps: Step J:Letxnew=x+ (x-x ) Step2: if(fb_ 2 fb- ), x-=x +1.2(x -x.),_l
w

14

13

1t

13

if (jf_

>

Jb_ )and (jf_ < jf_

), x = X+0.4(x - x)

where xw is the worst point in current simplex, fb is the fitness value, xnew is the new point, x,,, is the best point in current simplex, xmn is the worst point in current simplex. From Table 1, we can see that the fitness value of each Table 1 lists the fitness values of some new points new point in each new simplex is greater than the minimum calculated by simplex method in the first generation and the fitness value in each old simplex. This shows that proposed sixth generation. The first column represents the simplex

Fig. 3. Some tested images used in the experiments Image 1-3: bushes; Image 4-6: residential areas; Image 7-9: cultivated lands; Image 10-12: mountain lands; Image 13-15: dry lands.

1194

TABLE I
THE COMPUTATION RESULTS OF THE NEW POINTS IN THE SIMPLEX

No.

Max/Min

1st geneation
New point

Max/Min

6th generation New point 4.03 4.06 3.58 3.36 4.02 4.15 4.64 4.51 4.40 4.45

fbm=3.75
fb =,,,,2.75
2

3.74 3.52
2.98 3.17 3.37 3.18

fbm.=4.56
fbmm=3.52

fb-=3.75 fb ,,=1.86
fb,=3.75

fo,=4.56 fb,=3.52
fb-=4.56

fbm,m=2.33 fbm,,=2.35

fbm,,=3.78

fb.=3.75
5

2.43 2.40 4.48 4.44 TABLE II

fbm,,,-=4.56 fbmm=3.43
fb,,,=4.56

fbm, =3.75

fb,,=3.06

fb.==4.34

THE LIST OF THE 5 BESTFITNESS VALUES IN EACH GENERATION

Gen=71 3.75 3.70 3.29 3.21 3.19

Gen=2
4.44 4.43 3.75 3.73

3.70

Gen=3 4.44 4.34 4.27 4.05 3.99

Gen=4 4.44 4.34 4.33 4.27 4.12 TABLE m

Gen=5 4.56 4.51 4.44 4.34 4.34

Gen-6 4.56 4.52 4.51 4.46 4.44

Gen=7 4.56 4.52 4.52 4.51 4.51

THE COMPARIOSON RESULTS OF TEXTURE CLASSIFICATION RATE USING GA&GASPX

Method
GA GASPX

0.84 0.83

Residential area 0.82 0.86

Classification rates for different texture classes Cultivated Mountain


land 0.85 0.89 land 0.76 0.91

Dry land 0.81 0.82

Average

0.816 0.862

algorithm is able to ensure each new simplex become better and better with the increase of iteration times. This is helpful to prevent GA from the early maturation. Table 2 lists the five best fitness values of each generation from the first generation to the seventh generation. The table shows that the proposed approach tends to converge only after 7 generations ,and the average fitness value reaches 4.524. In contrast, traditional GA generally takes at least ten generations to reach such level. Table 3 shows the comparison results of classification rates on aerial images using GA and GASPX. From Table 3, it can be seen that the classification rates with the mask learned by GASPX are better than those with the mask learned by GA except bush textures. This illustrates that the

proposed method is able to obtain better solution than that obtained by GA. In other words, the proposed method has more powerful capability of classifying image textures.
V. CONCLUSIONS

In this paper we have shown how to integrate genetic algorithms and the simplex algorithm to generate the optimal mask used for texture classification. The hybrid learning approach based on genetic algorithms and the simplex algorithm was tested on real texture images. Experimental results are given to illustrate the capability of proposed method.

1195

Proper selection of control parameters for an application of GASPX is still an open issue. In this study we selected parameters manually. Further experiments with the adaptive parameter settings of the GASPX are necessary. In addition, the ability of proposed approach is limited by taining samples, and combination with other texture models and detection methods is not only helpfill, but sometimes is also necessary.
ACKNOWLEDGEMENT

The work was supported in part by the National Science Foundation ofChina under grant No. 40337055.
REFERENCES

[1] D.G Rogers, "A hybrid of friedman's multivariate adaptive regression splines (MARS) with Holland's genetic algorithm," Pro. of the 4th Int. Conf GeneticAlgorithms, San Diego CA ,pp,384-391,1991 [2] H Ishibuchi, N.Yamamoto, T. Murata and H Tanaka, "Genetic algorithms and neighborhood search algonthms for fizzy flowshop scheduling problems," Fuzzy Sets System, vol. 67, pp.81-100, 1994 [3] J.Yen, C .Liao, B. Lee and D. Randolph, "A hybrid approach to modeling metabolic systems using a genetic algorithms and simplex method," IEEE Thans. on Systems, Man, and Cybernetics-part B :cybernetics, Vol.28, pp.173-191, 1998 [4] P. S. de Souza and S. N. Talukdar, "Genetic algorithm in asynchronous teams," Prmc. 4th Int. Conf: Genetic Algorithms, San Diego, CA, pp.392-397, 1991 [5] K. E. Mathias, L. D. Whitley, C. Stork, and T. Kusuma, "Staged hybrid genetic search for seismic data imaging," Proc. 1st IEEE Conf Evolutionary Computation, Orlando, FL, pp. 356-361,1994 [6] Z Zheng, The Application of the Mathematic Programming in Surveying and Mapping, China Surveying and Mapping Press, 1993 [7] Z. Zheng and H Zheng, "Genetic algorithm for producing texture "tuned" masks," Pattern Recognition and Artficial Intelligence, Vol.14, pp.1 19-122, 2001

1196

Potrebbero piacerti anche