Optimum Multilevel Thresholding Hybrid Ga-pso by Algorithm

The conventional multilevel thresholding methods are efficient for bi-level thresholding. However, these methods are computationally very expensive for use in multilevel thresholding because the search of optimum threshold do in depth to optimize the objective function. To overcome these drawbacks, a hybrid method of Genetic Algorithm (GA) and Particle Swarm Optimization (PSO), called GA-PSO, based multilevel thresholding is presented in this paper. GA-PSO algorithm is used to find the optimal threshold value to maximize the objective function of the Otsu method. GA-PSO method proposed has been tested on five standard test images and compared with particle swarm optimization algorithm (PSO) and genetic algorithm (GA). The results showed the effectiveness in the search for optimal multilevel threshold of the proposed algorithm.


Introduction
In image understanding, image segmentation is always the first step of all steps of image processing and thresholding segmentation is one of the most common method used.In general, if the gray level image histogram is bi-model, the object image can be clearly distinguished from background.In this case, to choose the threshold value can be done easily by taking a value that is in the valley between the two histogram peaks.However, in the real world, the image gray level histogram is always multimodal.Therefore, it is not easy to determine appropriate threshold values in the histogram multimodal.Therefore, multilevel thresholding problem is considered as an important area of research among the research community, all over the world.
Over the years, many thresholding methods have been developed [1] [2].[3] has presented a comprehensive survey of various thresholding methods, among these methods, the global histogram-based algorithm is a widely used method for determining the threshold.Global thresholding methods can be classified into parametric and nonparametric approaches [4].In the parametric approach [5], the distribution of gray levels of each class has a probability density function that follows the Gaussian distribution.Parametric approach requires an expensive computational and time consuming.Nonparametric approach is used to determine the optimal threshold value based on certain criteria.Nonparametric approach is more accurate and reliable than the parametric method.
Otsu's method [6] chose the optimal threshold by maximizing the between-class variance of gray level.[7] found the optimal threshold value based on the maximization of the entropy of the histogram.[8] assumes that the gray level of each object in the image is normally distributed.Nonparametric approaches are simple and effective in bi-level thresholding.
Otsu and Kapur method can be easily extended to solve the problem of multilevel thresholding but not efficient in determining the optimal threshold because the computing time will increase exponentially.To improve efficiency, many methods have been proposed to solve the problem of multilevel thresholding [9].Liao, Chen, and Chung, (2001) have shown that the recursive algorithm successfully reduces the computational complexity in determining the multilevel threshold by accessing a look-up table method when compared with the conventional Otsu and Kapur.However, this method still has the problem that lasts a long time process when the number of threshold increases.
To overcome this problem, [10] apply GA in solving multilevel thresholding.Although the GA method has been successfully applied in solving linear optimization problems that complex, recent studies have found a decrease in efficiency in the performance of GA [11].The decrease in efficiency arises when the parameters are optimized to have high correlation.In addition, the premature convergence in GA can reduce the ability of search.
In this paper, the authors propose a method of optimization using GA-PSO with objective function of the Otsu method to solve the problem of multilevel thresholding in image segmentation.The proposed algorithm was tested on five standard test images and compared with the PSO and GA methods.
Particle swarm optimization (PSO) is one of the latest techniques for evolutionary optimization developed by Eberhart and Kennedy [12].PSO is based on the concept of a metaphor of social interaction such as bird flocking and fish schooling.
GA-PSO, first introduced by Kao and Zahara, a hybrid method that combines two heuristic algorithms, genetic algorithms (GA) and particle swarm optimization (PSO) [13].This hybrid method combines the concepts of GA and PSO by making individuals in a generation not just of the crossover and mutation as in GA, but also from the PSO mechanism.

Methodology
Otsu method is described as follows: assume that an image can be represented in a number of gray-level L (1, 2, ..., L).Number of pixels with gray level i denoted by i f then the total number of pixels equal to N (equal to).The possibility of a gray-level for an image is If an image can be divided into two classes, 0 C dan 1

Z Z
Using discriminant analysis, Otsu method is based on the between-class variance can be defined as follows: Where 2 ) 0 ( 0 0 For bi-level thresholding, Otsu choose the optimal threshold that maximizes the betweenclass variance , 2 B V so that: Above formula can be easily extended to solve multi-level thresholding on an image.Assume that there are a number of m-1 threshold (t 1 ,t 2 ,...t m-1 ), that the original image is divided into a number of class m : 0 C for [1,..
Otsu and Kapur method has been proven as an efficient method to solve bi-level thresholding problems in image segmentation.However, when this method is extended to multilevel thresholding, computing time increases exponentially with increasing number of threshold.To overcome this problem, this paper offers a method of Kapur and Otsu optimization based on GA-PSO to overcome the problem of multilevel thresholding.
Application of GA-PSO with PSO Otsu algorithm using two random variables r1 and r2 are both generating random numbers with a range between 0 and 1. Random variables are used to give effect to the stochastic nature of the PSO algorithm.The value of r1 and r2 will be adjusted by a constant c1 and c2 that has a range of values between 0 <c1, c2 " 2. The Constants called acceleration coefficients which affect the maximum distance that can be taken by a particle in an iteration.Update the velocity of a particle is distinguished for each dimension j ¼ ... n (n is based on the number of optimized parameters), so that vi,, j j represents the dimension of the velocity vector associated with the i-th particle.So the equation for the velocity update by van den Bergh (2001) can be defined as follows: vi=…vi + c1r1 (pbest-xi) + c2r2 (gbest-xi) (8) where, vi = velocity of the particle current … = weight of inertia, c1, c2 = acceleration coefficient r1, r2 = uniform random numbers between 0 and 1 pbest = best position particle current gbest = the global position of the current best particle xi = the particle position at this time It can be seen from the velocity update that equation c2 set a maximum distance that a particle is affected by the global best, and c1 adjust the distance that is affected by the personal best position of the particle.The value of vi can be limited to the value [-VMAX, VMAX] to prevent the occurrence of events in which the particles leave the search area.If the search area is restricted to [-xmax, xmax], then the value of VMAX is usually defined as the VMAX, k xmax, where 0.1 " N " 1.0.The position of each particle is updated using the following equation so that the resulting equation: where, xi = the new particle positions x = position of current particle vi = new particle velocity From the theory of GA and PSO it found that each of them can be combined to obtain a better optimization results.The first step to solve that problem is to do a first coefficient of the GA to find the individual process.After that, the PSO will perform the GA Tracking finding the best individual.In the application of the PSO velocity on each individual do tracking and weighting based on the value of fitness.So after the GA and PSO is given, it will generate a sorted best individual proposals.With the value of the velocity, the result will have more individual choice and there are also traces of the previous individual.

Results and Analysis
In order to evaluate the performance of the proposed method, our algorithm has been tested using images 1-5 from Matlab standard images.figure 1 show on of the graphic histogram after we running the code from Matlab.We have compared our method with others algorithm : (i) GA-Otsu thresholding method, (ii) PSO-Otsu thresholding method.The compare methods shown in table I.
In this research trials conducted in Matlab.Matlab provides a standard image that is ready to do the tests.Test images will be used as the matrix first and then will see a histogram and then analyzed the results of thresholding.The tables tell us compare methods that we are already tested it.Number 1-4 tell the rank for fitness function.The fitness function tell the multilevel threshold.GA, PSO, and GA-PSO are methods compared to find multilevel threshold.GA-PSO is method that proposed in this paper.
At the trial is conducted experiments on a standard image Matlab.This standard image histogram and the search process is carried out with a threshold value of the genetic algorithm in advance.Once the threshold is found using genetic algorithms, for further analysis using PSO.To generate value multithreshold.In the above data are presented in the effectiveness of the search threshold value is based on the value of fitness.

Conclusion
In table I above it can be seen that the GA-PSO algorithm can have a value higher than the effectiveness of both algorithms GA or PSO.These results may indicate that the hybrid algorithm can be better in the search multithresholding value.In the next study you can replace the value of fitness not only of Otsu, but also the formulation of other thresholding.From that table, can tell us that fitness value that our proposed method is bigger than others.It means our proposed method is efficient and find good multilevel threshold than two others.
If T P is the mean intensity for the entire image,

Fig 1 .
Fig 1. Image and the histogram from code.