Page 47 - Fister jr., Iztok, and Andrej Brodnik (eds.). StuCoSReC. Proceedings of the 2015 2nd Student Computer Science Research Conference. Koper: University of Primorska Press, 2015
P. 47
parison of DE strategies for Gray-Level MultiLevel
Thresholding
Uroš Mlakar Iztok Fister
University of Maribor University of Maribor
Smetanova 17 Smetanova 17
Maribor, Slovenia Maribor, Slovenia
uros.mlakar@um.si iztok.fister@um.si
ABSTRACT class variance or various entropy measures. Many heuristic
methods have gained a lot of attention recently, since ex-
In this paper we investigate the impact of different mutation haustive methods are usually computationally inefficient.
strategies in differential evolution applied to the problem
of gray-level multilevel thresholding. Four different strate- In this paper we investigate the influence of different muta-
gies are compared, namely rand/1, best/1, rand to best/1, tion strategies on the quality of segmentation by using the
and current to best/1. These strategies are tested on four between-class variance as the objective function proposed by
standard test images taken from literature. The quality of Otsu [5]. The rest of the paper is organized as follows. In
segmented images was compared on the basis of the PSNR Section 2 some of related work is presented in the area of
metric, which showed that the best performing strategy was multilevel thresholding. In Section 3, image segmentation
current to best/1 based on the mean PSNR value, but in and the Otsu crietrion are described, while in Section 4 the
the case of best result found, the rand/1 strategy performed differential evolution algorithm (DE), along with 4 muta-
best. When comparing the mean and best objective values, tion strategies are presented, which have been used for this
the best/1 strategy outperformed the others. study. In Section 5 we define the metric, which was used to
assess the quality of the segmentation and present the ex-
Keywords perimental results. In Section 6 we will conclude this paper
with some findings.
Multilevel thresholding, Otsu criterion, evolutionary algo-
rithm, differential evolution 2. RELATED WORK
1. INTRODUCTION Many works has been done for multilevel thresholding using
evolutionary algorithms. Some of these algorithms can be
Image segmentation is a process of dividing an image into found in [7] and [4]. Alihodzic et al. [1] introduced a hybrid
disjoint sets, which share similar properties such as intensity bat algorithm with elements from the differential evolution
or color. Image segmentation is usually the first step for and artificial bee colony algorithms. Their result show their
many high-level methods, such as feature extraction, image algorithm outperforms all other in the study, while signifi-
recognition, and classification of objects. Simply put, image cantly improving the convergence speed. Bhandari et al. [2]
segmentation is a process of dividing an image into regions, investigated the suitability of the cuckoo search (CS) and the
which are used as input for further specific applications. Im- wind driven optimization algorithm for multilevel threshold-
age thresholding is one of the most used and simplest seg- ing using Kapur’s entropy as the objective function. The
mentation techniques, which performs image segmentation algorithms were tested on a standard set of satellite images
based on values contained in the image histogram. In the by using various number of thresholds. They concluded that
case of separating an image into two classes, the process is both algorithms can be efficiently used for the multilevel
called bilevel thresholding, but when separating the image thresholding problem. Duraisamy et al. [3] proposed a novel
into several regions we deal with multilevel thresholding. particle swarm optimization (PSO) algorithm maximizing
The selection of optimal threshold values is crucial, since the Kapur’s entropy and the between-class variance. Their
the results of good segmentation are a good foundation for algorithm has been tested on 10 images, and the results
applications, which further process the segmented images. compared with a genetic algorithm. The PSO proved better
in terms of solution quality, convergence, and robustness.
Multilevel thresholding can be regarded as an optimization Zhang et al. [8] presented an artificial bee colony algorithm
process, usually maximizing certain criteria like between- (ABC), for the problem of multilevel thresholding. They
compared their algorithm to PSO and GA, where their con-
clusions were that the ABC is more rapid and effective, using
the Tsallis entropy.
3. IMAGE SEGMENTATION
For the purpose of multilevel threshold selection the Otsu
criterion was selected as the objective function. It operates
on the histogram of the image, maximizing the between-class
StuCoSReC Proceedings of the 2015 2nd Student Computer Science Research Conference 47
Ljubljana, Slovenia, 6 October
Thresholding
Uroš Mlakar Iztok Fister
University of Maribor University of Maribor
Smetanova 17 Smetanova 17
Maribor, Slovenia Maribor, Slovenia
uros.mlakar@um.si iztok.fister@um.si
ABSTRACT class variance or various entropy measures. Many heuristic
methods have gained a lot of attention recently, since ex-
In this paper we investigate the impact of different mutation haustive methods are usually computationally inefficient.
strategies in differential evolution applied to the problem
of gray-level multilevel thresholding. Four different strate- In this paper we investigate the influence of different muta-
gies are compared, namely rand/1, best/1, rand to best/1, tion strategies on the quality of segmentation by using the
and current to best/1. These strategies are tested on four between-class variance as the objective function proposed by
standard test images taken from literature. The quality of Otsu [5]. The rest of the paper is organized as follows. In
segmented images was compared on the basis of the PSNR Section 2 some of related work is presented in the area of
metric, which showed that the best performing strategy was multilevel thresholding. In Section 3, image segmentation
current to best/1 based on the mean PSNR value, but in and the Otsu crietrion are described, while in Section 4 the
the case of best result found, the rand/1 strategy performed differential evolution algorithm (DE), along with 4 muta-
best. When comparing the mean and best objective values, tion strategies are presented, which have been used for this
the best/1 strategy outperformed the others. study. In Section 5 we define the metric, which was used to
assess the quality of the segmentation and present the ex-
Keywords perimental results. In Section 6 we will conclude this paper
with some findings.
Multilevel thresholding, Otsu criterion, evolutionary algo-
rithm, differential evolution 2. RELATED WORK
1. INTRODUCTION Many works has been done for multilevel thresholding using
evolutionary algorithms. Some of these algorithms can be
Image segmentation is a process of dividing an image into found in [7] and [4]. Alihodzic et al. [1] introduced a hybrid
disjoint sets, which share similar properties such as intensity bat algorithm with elements from the differential evolution
or color. Image segmentation is usually the first step for and artificial bee colony algorithms. Their result show their
many high-level methods, such as feature extraction, image algorithm outperforms all other in the study, while signifi-
recognition, and classification of objects. Simply put, image cantly improving the convergence speed. Bhandari et al. [2]
segmentation is a process of dividing an image into regions, investigated the suitability of the cuckoo search (CS) and the
which are used as input for further specific applications. Im- wind driven optimization algorithm for multilevel threshold-
age thresholding is one of the most used and simplest seg- ing using Kapur’s entropy as the objective function. The
mentation techniques, which performs image segmentation algorithms were tested on a standard set of satellite images
based on values contained in the image histogram. In the by using various number of thresholds. They concluded that
case of separating an image into two classes, the process is both algorithms can be efficiently used for the multilevel
called bilevel thresholding, but when separating the image thresholding problem. Duraisamy et al. [3] proposed a novel
into several regions we deal with multilevel thresholding. particle swarm optimization (PSO) algorithm maximizing
The selection of optimal threshold values is crucial, since the Kapur’s entropy and the between-class variance. Their
the results of good segmentation are a good foundation for algorithm has been tested on 10 images, and the results
applications, which further process the segmented images. compared with a genetic algorithm. The PSO proved better
in terms of solution quality, convergence, and robustness.
Multilevel thresholding can be regarded as an optimization Zhang et al. [8] presented an artificial bee colony algorithm
process, usually maximizing certain criteria like between- (ABC), for the problem of multilevel thresholding. They
compared their algorithm to PSO and GA, where their con-
clusions were that the ABC is more rapid and effective, using
the Tsallis entropy.
3. IMAGE SEGMENTATION
For the purpose of multilevel threshold selection the Otsu
criterion was selected as the objective function. It operates
on the histogram of the image, maximizing the between-class
StuCoSReC Proceedings of the 2015 2nd Student Computer Science Research Conference 47
Ljubljana, Slovenia, 6 October
