Vous êtes sur la page 1sur 6

Multi-objective Nature-inspired Clustering

Techniques for Image Segmentation


Bong Chin Wei Rajeswari Mandava
School of Computer Science, School of Computer Science,
Universiti Sains Malaysia, Universiti Sains Malaysia,
Penang, Malaysia. Penang, Malaysia.
bongwendy@gmail.com or cwbong@cs.usm.my mandava@cs.usm.my

Abstract Image segmentation aims to partition an image homogeneity, spatial coherence, connectivity, and so on.
into several disjointed regions that are homogeneous with regards Solution design with consideration of multiple objectives
to some measures so that subsequent higher level computer vision combination is difficult problem, thus creating a gap between
processing, such as object recognition, image understanding and the nature of image segmentation problem and real-world
scene description can be performed. Multi-objective practical solution. Multi-objective nature-inspired clustering
formulations are realistic models for image segmentation because (MoNiC) approach is an appropriate method to bridge this gap
objectives under consideration conflict with each other, and [13, 14].
optimizing a particular solution with respect to a single objective The purpose of this study is to propose a generic design and
can result in unacceptable results with respect to the other
modelling of the MoNiC method for applications of image
objectives. In this paper, we present the current multi-objective
nature-inspired clustering (MoNiC) techniques for image
segmentation. First, we explain how image segmentation is a
segmentation. We are able to diagnose the requirements and problem with multiple objectives. In this section, we identify
issues for modelling this specific technique in the image multiple objectives associated with image segmentation
segmentation problem. Three identified important phases include problem. We elaborate on the characteristics of multi-objective
intelligence, design and choice with respect to the issues of optimization model critical for the image segmentation
clustering problem of image segmentation and multi-objective decisions in Section 3. In Section 4, we describe the current
clustering algorithm design. MoNiC methods that has been applied in image segmentation
problem. At last, the paper presents the modelling and design
Keywordsclustering, image processing, nature-inspired issues of MoNiC in application of image segmentation before
techniques we conclude the survey.
I. INTRODUCTION II. IMAGE SEGMENTATION - A MULTI-OBJECTIVE
Image segmentation has long been an important and OPTIMIZATION PROBLEM
challenging topic in the field of digital image processing. It Image segmentation problem is to decompose a given image
aims to partition an image into several disjointed regions that into segments, or sometimes called regions containing similar
are homogeneous with regards to some measures so that pixels. Example of segments might be regions of the image
subsequent higher level computer vision processing, such as depicting the same object. It is typically defined as an
object recognition, image understanding and scene description exhaustive partitioning of an input image into regions, each of
can be performed [1, 2]. The image segmentation process is which is considered to be homogeneous with respect to some
considered as an essential component of any image analysis image property of interest (e.g., intensity, color, or texture)
system and this problem has received a great deal of attention. [15]. This basic idea, depicted in Figure 1, has been
It is also considered as the most difficult low-level task successfully used for intensity images (with or without
because the segmentation performance needs to be adapted to texture), range (depth) images and multispectral images.
the changes in image quality. It is affected by variations in
environmental conditions, imaging devices, time of day, and (a)
so on [3, 4]. Despite the large number of segmentation
techniques presently available, no general method has been
found that perform adequately across a diverse set of imagery
[3-5].
A new trend of problem formulation for image
segmentation is to use approaches based on multiple Image Image
objectives in its decision making process [6-12]. For problems data Feature Space
with multiple objectives, the objective functions defined are
generally conflicting, preventing simultaneous optimization of Figure 1. Image measurements and positions are transformed to features.
each objective. Many, or even most, real image segmentation Clusters in feature space correspond to image segments
problems actually do have multiple objectives, such as feature

978-1-4244-6502-6/10/$26.00 2010
c IEEE 150
assessment). There are possibilities of multiple sources of
information for a segmentation problem, thus multiple
Image pattern Goal Features for consideration dimensions or multiple representations have to be considered.
- Intensity
In the segmentation process, there is also the favour of
- Colour
Feature selection - Shape combination of multiple methods in getting the appropriate
/extraction - Texture output. With ensemble of multiple methods, there is a
- Spatial tendency of multiple optimizations and decision making
Pattern Feedback
representation loop
processes where multiple validity assessments should be used.

Inter-pattern similarity III. APPROACHES BASED MULTIPLE OBJECTIVES (MO)


Criteria for consideration
Optimization & - Spatial coherence vs.
There are two general approaches of MO [16, 17]. The first
decision making Feature homogeneity approach is to combine the multiple objective functions into a
Grouping - Connectedness vs. single composite function or move all but one objective to the
compactness constraint set. The second general approach is to determine a
Number of Cluster, c - Diversity vs. accuracy set of solutions that are non-dominated with respect to each
objective. The following discussion shows the formulation of
Validity assessment
the two approaches.
A. The Conventional Weighted-Formula Approach
Figure 2. Stages in image segmentation and the consideration of multiple (WFA)
objective criteria in each stage.
The most commonly used approach is to transform a MO
Feature selection is the process of identifying the most problem into a problem with single objective function. This is
effective subset of the original features to use in clustering. typically done by assigning a numerical weight to each
Feature extraction is the use of one or more transformations of objective (evaluation criterion) and then combining the values
the input features to produce new salient features. Either or of the weighted criteria into a single value by either adding or
both of these techniques can be used to obtain an appropriate multiplying all the weighted criteria. That is, the quality Q of a
set of features. For instance, in segmenting a medical image given candidate model is typically given by one of the two
based on CT scan, multiple features related intensity, shape kinds of formula:
and spatial relationship could be considered. Q = w1 c1 + w2 c2 + . . . + wn cn , (1) or
Pattern proximity is usually measured by a distance Q = c1W1 c2W2 . . . cnWn (2)
function defined on pairs of patterns. A simple distance where wi, i = 1,n, denotes the weight assigned to criteria ci,
measure like Euclidean distance can often be used to reflect and n is the number of evaluation criteria.
the dissimilarity between two patterns, whereas other B. Pareto approach (PTA)
similarity measures can be used to characterize the conceptual
similarity between patterns. Consideration of the inter-pattern The second approach is called the Pareto approach (PTA). The
similarity is an issue related to multiple criteria. For example, basic idea of the Pareto approach is that, instead of
segmenting a medical image based on maximization of spatial transforming a problem of MO into a single-objective function
coherence and feature homogeneity. Another possible pair of and then solving it by using a single-objective search method,
objective criteria includes maximize inter cluster distances (i.e. one uses a multi-objective algorithm to solve the problem. The
increase separation) and minimize intra-cluster distances (i.e. formulation starts with the simultaneous optimization of
increase compactness). several objectives. Mathematically, the problem can be written
The grouping step can be performed in a number of ways. as follows [18, 19],
The output result can be hard (a partition of the data into Minimize
G G G G G
groups) or fuzzy (where each pattern has a variable degree of f ( x ) = [ f1 ( x ), f 2 ( x ),..., f k ( x )]T (3)
membership in each of the output clusters). Different Subject to
algorithms produce different partitions/groups based on a set G
gi ( x ) 0, i = 1,2,..., m (4)
of selected criteria. A feedback path from the output of the G
grouping process could affect subsequent feature extraction h j ( x ) 0, j = 1,2,..., p (5)
and similarity computations. Cluster validity analysis is the G
where x = [ x i , x2 ,..., xn ]T is the vector of decision variables,
assessment of a clustering procedures output. This analysis
often uses a specific criterion of optimality but is usually fi : n , i = 1,2,..., k are the objective functions and
arrived at subjectively. Hence, gold standards rarely exist in
clustering except in well-prescribed sub-domains. g i , h j : n , i = 1,2,..., m, j = 1,2,..., p are the constraint
In short, image segmentation is a multi-objective functions of the problem.
optimization problem. The consideration of multiple criteria A reasonable solution is to investigate a set of solutions,
(objectives) starts from the understanding the data point of each of which satisfies the objectives at an acceptable level
view (image pattern) to its selected segmentation process without being dominated by any other solution. Those
involved (feature selection/extraction, similarity/ dissimilarity solutions are called non-inferior or non-dominated solutions
measure) and finally the assessment of its output (validity and the region of those solutions is called Pareto front [17-19].

2010 IEEE Conference on Cybernetics and Intelligent Systems 151


The first goal is to focus (intensifying) the optimization search which is stochastic technique. Evolutionary approaches,
on a particular region of the Pareto front. In Figure 3, the motivated by natural evolution, make use of evolutionary
region of points defined by a bold line is used to mark the operators and a population of solutions to obtain the globally
Pareto front of a bi-objective problem.The second goal optimal partition of the data [13]. The popular variants of the
demands the search effort to be uniformly distributed over the evolutionary algorithm used include NSGA-II [9, 10, 26],
Pareto front. The third goal aims at extending the Pareto front PESA-II [6, 27, 28] and SPEA-II [29] (refer Table 1). The
at both ends, exploring new extreme solutions. details of the step-by-step algorithmic design for each of these
heuristic methods with evolutionary/genetic algorithm and its
variants may be found in [17]. Other nature-inspired heuristics
f1 are the simulated annealing approach (SA) [7, 20, 21],
differential evolution [8, 11] and ant colony optimization [30].

TABLE I. APPLICATIONS OF MONIC TECHNIQUES IN IMAGE


F SEGMENTATION

Type of
No. of Data- Type of
nature-
WFA/ PTA Ref objective sets repre-
f2 inspired
functions used* sentation
technique
NSGA-II PTA [9, 31] R Centroid
Figure 3. An example of a problem with two objective functions The Pareto NSGA-II WFA [26] R, M Centroid
front is marked with a bold line PESA-II PTA [28] S Graph
PESA-II WFA [6, 27] N Label
IV. MULTI-OBJECTIVE NATURE-INSPIRED CLUSTERING SPEA-II PTA [29] S Graph
(MONIC) TECHNIQUES Evolutionary
PTA [32] S Graph
method
In order to compare the current applications of MoNiC, Evolutionary M, N,
PTA [33] Centroid
Table 1 provides a brief summary. The applications of these method S
Simulated [7, 20, TWO R, M,
clustering methods have been classified into the two common PTA Centroid
annealing 21] N, S
approaches: WFA or PTA (refer section III). Table 1 also Differential M, N,
shows the type of nature-inspired method used, the number of PTA [8] Centroid
Evolution S
objectives considered, the datasets tested (natural, medical, Differential
PTA [11] R Centroid
remote sensing or simulated images) and the definition of the Evolution
Ant Colony
type of representation methods (centroid-based, graph-based Optimization
PTA [30] N Centroid
or label-based). Immune
PTA [34] N, S Centroid
The common objectives functions used in MO clustering Algorithm
techniques are related to cluster validity measure which NSGA-II PTA [10] M Centroid
usually produces an index value [9, 10, 20, 21]. The role of Particle [12]
THREE
Swam PTA R -
these validity indices is very important in determining the Optimization
appropriate number of clusters presented in a dataset. The * M - Medical, N - Natural, R - Remote Sensing and S - Simulated or Handcrafted
optimization process aims for optimal number of clusters with
optimal clustering output. Among the large number of cluster TABLE II. SOME OF THE POST-PROCESSING METHODS IN MONIC FOR
IMAGE SEGMENTATION PROBLEM
validity indices, the DaviesBouldin (DB) index and Dunns
index are highly used for crisp clustering, whereas, the Xie Method Description
Beni (XB) index is widely used for fuzzy clustering [22]. /index
Gap Gap Statistics uses reference distribution in the process of
Indices used are discussed in detail in Jain and Dubes [23] and statistic estimating the number of clusters. This technique compares
are not discussed further in this paper. However, none of the [28, 32] the within-cluster dispersion with what one might expect
indices perform satisfactorily for wide range of data sets [15, given a reference null distribution.
24]. Most of the clustering validation techniques are biased Adjusted The adjusted Rand index assumes the generalized hyper-
Rand geometric distribution as the model of randomness. It is
towards a clustering criterion [25]. Therefore, optimizing two Index [28] index of choice in many researches as measure of
to three cluster validity functions is important because they agreement between the external criteria and clustering
can complement and compensate one another. results
The optimization/search techniques used in MO techniques Support A support vector machine constructs two hyper-planes (a
Vector generalization of the concept of a plane into a different
are divided into deterministic and stochastic search techniques
Machine number of dimensions) in a high-dimensional space. A
[15]. Deterministic search techniques guarantee an optimal [9] good separation is achieved by the hyper-plane that has the
partition by performing exhaustive enumeration. Meanwhile, largest distance to the nearest training data-points of any
the stochastic search techniques generate a near-optimal class.
partition reasonably quickly, and guarantee convergence to Fuzzy Allows a fuzzy decision between several results of a cluster
voting algorithm. It is done by identifying the points assigned to
optimal partition asymptotically. Most of the nature-inspired technique certain clusters with high fuzzy membership by most of the
search techniques used is related to evolutionary approach, [9] non-dominated solutions.

152 2010 IEEE Conference on Cybernetics and Intelligent Systems


Several mechanisms have been used for the post- Another challenge to the more successful use of clustering
processing of filtering the Pareto-optimal solutions to get a includes the high computational complexity of many
single best result in Pareto approach. They include Support clustering algorithms. Clustering is subjective in nature. A
Vector Machine [9], goodness/validity measure [7, 20, 21], same data set needs to be partitioned differently to serve
Gap statistic [29] and so on. In [28], a post-processing with different purposes [15]. Approaches based on genetic
Model Selection Testing Procedure has been tested with the algorithms and simulated annealing are optimization
Gap statistic or the attainment score method to select the techniques that are restricted to reasonably small data sets.
estimated best solution from the Adjusted Rand Index value. Implementations of conceptual clustering optimize some
Each of these post-processing methods is briefly explained in criterion functions and are typically computationally
Table 2. Besides the listed methods in the Table 2, other expensive [12]. Thus, extension of the current research is
validity measurement indices may be used for similar reason. needed for the improvement of the convergence speed of an
algorithm. The approaches should permit a substantial
V. MODELLING AND DESIGN ISSUES reduction in the number of total evaluations required [38]. A
To solve image segmentation problem with MoNiC, one scalable framework requires both cluster ensemble creation
should consider two main streams of design and and merging to be efficient in terms of time and memory
implementation issues: the nature of clustering problem in complexity [19, 39]. Future work should be conducted to
image segmentation problem (refer as CPIS) and the nature of handle different numbers of clusters in different clustering
multi-objective clustering algorithm (refer as MOCA). A solutions, experiments on large data sets, and automatically
successful design should start with understanding of their detecting the threshold value in centroid filtering.
strengths and weaknesses and how image segmentation can be Therefore, the criteria for good segmentation are [4], (1) the
translated into MoNiC terms [14]. A thorough investigation of segmented regions should uniform and homogeneous with
the current research experience will benefit those who are respect to some characteristic, such as gray value or texture, (2)
interested in this research arena. region interiors should be free of holes and region boundaries
should be smooth and spatially accurate, and (3) adjacent
A. Clustering Problem for Image Segmentation (CPIS)
regions should be differing significantly based on the
When presented with a new image, selecting the appropriate characteristic on which they are uniform. If one represents this
set of algorithm parameters is the key to effectively criteria set in terms of a function, then the problem of good
segmenting the image [4]. In reality, there exist several factors segmentation is optimizing this objective function by selecting
which make the parameter adjustment process very difficult appropriate segmentation parameters. Combining multiple
[3]. First, the number of parameters present in a typical clustering solutions address the issue of scalability, the
clustering algorithm is usually quite large. Second, the distributed nature of some data, and the robustness or stability
parameters mutually interact in a complex, non-linear fashion, of the clustering solution [40].
which makes it difficult or impossible to model their behavior B. Multi-objective Clustering Algorithm (MOCA)
in an algorithmic or rule-based fashion. Third, since variations
between images cause changes in the clustering results, the As the formulation of MOCA can yield various results, the use
objective function that represents the clustering quality also of suitable model should be considered from many aspects and
varies from image to image. Finally, the definition of the the main purpose of the application. The situation in
objective function itself can be a subject of debate because conventional weighted formula is comparatively simpler [41].
there is no single, universally accepted measure of However, the problem of this method lies in the proper
performance available with which to uniquely define the selection of the weights or utility functions to characterize the
quality of the clustered image. decision-makers preferences. In practice, it can be very
Most real images are rich in color and texture features, difficult to precisely and accurately select these weights, even
and this fact makes it very difficult to recognize objects in an for someone familiar with the problem domain [16, 17, 42].
image accurately. Two typical problems in image Compounding this drawback is that scaling amongst objectives
segmentation process are [35]: (1) over-segmentation: an is needed and small perturbations in the weights can
object is partitioned into multiple regions after the sometimes lead to quite different solutions.
segmentation; and (2) under-segmentation: multiple objects Selecting the best solution(s) from the Pareto-optimal front
are represented by a single region after segmentation. Over- is important in multi-objective clustering but it is difficult
segmentation usually occurs when images contain complicated problem [42]. An unfortunate initialization and the usual noise
and very detailed local textures [36, 37]. In these cases, exist in any data set and they can cause a poor result.
traditional approaches normally cannot produce satisfactory Currently, although there are some semi-supervised methods
results because over-segmentation produces minute segments of selecting a single solution from the Pareto-optimal front,
that confuse the global information and prevent further these methods assume that the labelling of the partial data
analysis [36]. Over-segmentation also produces erroneous points is known before hand. Besides, the present of image
results in noisy images. It is well known that some modalities noise problem is always been ignored.
of remote sensing images, such as SAR images, are often Most of the publications deal with two- or three-objective
polluted by speckle noise. The presence of noise greatly problems (Table 1). Studies addressing high-dimensional
affects clustering results using traditional approaches. problems (more than three objectives) are rare. A high number

2010 IEEE Conference on Cybernetics and Intelligent Systems 153


of objectives cause additional challenges [43, 44]. The main
challenge may be due to the proportion of non-dominated Problem definition
solutions in a population increases rapidly with the number of - Formulation of objective
functions Intelligence phase
objectives [44, 45]. When there are more than three objectives, - Evaluation criteria (CPIS)
recent studies have shown that multi-objective evolutionary - Constraints
algorithm based on Pareto-optimality have difficulties to find a - Parameter initialization
good Pareto front approximation [44-46]. Thus, an interesting - Data representation
future work is to study the conflict analysis among objectives
Algorithmic optimization
and the influence of a higher (or lower) number of objectives. - Alternatives Design phase
Other interesting topic includes the study of scalable - Search directions (MOCA)
framework that requires an algorithm to be efficient in terms - Decision-makers involvements
of time and memory complexity [39, 40, 47]. - Parameter tuning
- Time complexity
Use of other potential nature-inspired methods is remained
to be explored for MoNiC in image segmentation. While many
effective algorithms have been developed for clustering, no
Evaluation Choice Phase
single algorithm has been shown to be either empirically or
Sensitivity analysis (CPIS/MOCA)
theoretically better than other algorithms in all scenarios [48]. Recommendations
There is a need to utilize other recent MOCA techniques like
particle swarm optimization, artificial immune systems, ant- Figure 4. Framework of multi-objective clustering for image segmentation
colony optimization, scatter search and so on [19, 49].
REFERENCES
VI. SUMMARY AND CONCLUSION [1] Gonzalez, R.C., and Woods, R.E., Digital Image processing. 3 ed. 2007:
The development of a new MOP method should consider Prentice Hall.
factors such as robustness, interactivity, generality, [2] Xia, Y., Feng, D.D., Wang, T., Zhao, R. and Zhang, Y., Image
Segmentation by clustering of spatial patterns. Pattern Recognition
simplification, and consistency [5, 50-52]. To summarize the Letters, 2007. 28: p. 1548-1555.
key issues of the design and development of multi-objective [3] Bhanu, B., Lee, S. and Das, S., Adaptive image segmentation using
clustering problem for image segmentation, Figure 3 presents multi-objective evaluation and hybrid search methods. Machine
three important phases that involve intelligence, design and Learning in Computer Vision, 1993.
choice. The problem definition overlaps the intelligence phase [4] Bhanu, B., Lee, S. and Das, S., Adaptive image segmentation using
of decision making. It involves searching the decision genetic and hybrid search methods. IEEE Trans. on Aerospace and
environment for conditions calling for decisions; raw data are Electronic Systems, 1995. 31(4): p. 1268 - 1291.
obtained, processed and examined for clues that may identify [5] Zhang, Y.J., A Survey on Evaluation Methods for Image Segmentation.
Pattern Recognition, 1996. 29(8): p. 1335-1346.
opportunities for problems. The CPIS issues should be included
in this phase. The design phase involves inventing, developing [6] Shirakawa, S., and Nagao, T. Evolutionary Image Segmentation Based
on Multiobjective Clustering. in Proc. of Congress on Evolutionary
and analyzing a set of possible solutions or alternative course Computation (CEC '09). 2009. Trondheim, Norway.
of actions in terms of algorithm optimization to problem [7] Saha, S., and Bandyopadhyay, S., A new symmetry based
identified in the intelligence phase. The generation of multiobjective clustering technique for automatic evolution of clusters.
alternative decisions is purely part of the design stage. The Pattern Recognition, 2010. 43(4): p. 738-751.
issues of MOCA should be taken care of. The evaluation of [8] Swagatam, D., Ajith, A., and Amit, K., Clustering Using Multi-
alternative is mainly part of the choice phase. The choice is objective Differential Evolution Algorithms, in Chapter 6 of
what many people think of as making a decision. It involves Metaheuristic Clustering. 2009.
selecting a particular alternative from those available. At this [9] Mukhopadhyay, A., and Maulik, U., Unsupervised pixel classification in
satellite imagery using multiobjective fuzzy clustering combined with
phase, each alternative is evaluated and analyzed in relation to SVM classifier. IEEE Trans. on Geoscience and Remote Sensing 2009a.
others in terms of a specified rule. The rule is used to rank the 47 (4).
alternatives under consideration. The ranking depends on the [10] Mukhopadhyay, A., Maulik, U., and Bandyopadhyay, S. Multiobjective
decision makers preferences and it can be acquired before or genetic clustering with ensemble among pareto front solutions:
during the optimization process. Thus, both the CPIS and Application to MRI brain image segmentation. in Proc. of 7th
MOCA issues are to be noted here. The three stages of decision International Conference on Advances in Pattern Recognition. 2009b.
making do not necessarily follow a linear path from [11] Saha, I., Maulik, U. and Bandyopadhyay, S. An Improved Multi-
intelligence, to design to choice. At any point in the decision objective Technique for Fuzzy Clustering with Application to IRS Image
Segmentation. in Proc. of EvoWorkshops 2009 on Applications of
making process, it may be necessary to loop back to an earlier Evolutionary Computing. 2009.
phase. In short, each stage of decision-making process requires [12] Paoli, A., Melgani, F., and Pasolli, E., Clustering of Hyperspectral
different types of information. Images Based on Multiobjective Particle Swarm Optimization. IEEE
Trans. on Geoscience and Remote Sensing, 2009 (in press).
ACKNOWLEDGMENT [13] Guliashki, V., Toshev, H., and Korsemov,C., Survey of Evolutionary
This work was supported by Universiti Sains Malaysia Algorithms used in multiobjective optimization. Problems of
Engineering Cybernetics and Robotics, Bulgarian Academy of Sciences,
Research University Grant under grant number 2009.
1001/PKOMP/817001

154 2010 IEEE Conference on Cybernetics and Intelligent Systems


[14] Jones, D.F., Mirrazavi, S. K. and Tamiz, M., Multi-objective meta- objective clustering approaches. in Proc. of 10th Brazilian Symposium
heuristics: An overview of the current state-of-the-art. European on Neural Networks, SBRN 2008. 2008.
Journal of Operational Research, 2002. 137(1): p. 1-9. [34] Gong, M., Zhang, L., Jiao, L., and Gou, S. Solving multiobjective
[15] Jain, A.K., Murty, M.N., and Flynn, P.J., Data Clustering: A Review. clustering using an immune-inspired algorithm. in Proc. of IEEE
ACM Computing Surveys, 1999. 31(3). Congress on Evolutionary Computation. 2007.
[16] Savic, D. Single objective vs. Multiple objectives Optimisation for [35] Zhuo, C., Chin, F.Y.L., and Chung, R.H.Y., Automated Hierarchical
Integrated Decision. in Proc. of The First Biennial Meeting of the Image Segmentation Based on Merging of Quadrilaterals. WSEAS
International Environmental Modelling and Software Society 2002. Transactions on Signal Processing, 2006. 2(8): p. 1063-1068.
[17] Konak, A., Coit, D.W., and Smith, A.E., Multi-objective optimization [36] Zheng, L., Chan, A. and Liu, J.S. DWT based MMRF segmentation
using genetic algorithms: A tutorial. Reliability Engineering and algorithm for remote sensing image processing. in Proc. of International
System Safety, 2006. 91: p. 9921007. Geoscience and Remote Sensing Symposium. 1999.
[18] Coello, C.A.C., A Comprehensive Survey of Evolutionary-Based [37] Chevrefils, C., Cheriet, F., Aubin, C.-E., and and Grimard, G., Texture
Multiobjective Optimization Techniques. Knowledge and Information Analysis for Automatic Segmentation of Intervertebral Disks of
Systems, 1998. 1(3): p. 129-156. Scoliotic Spines From MR Images. IEEE Trans. on Information
[19] Coello, C.A.C., Evolutionary multi-objective optimization: some Technology in Biomedicine, 2009. 13(4): p. 608 - 620.
current research trends and topics that remain to be explored. Frontiers [38] Handl, J., and Knowles, J. Improvements to the scalability of
of Computer Science in China 2009. 3(1): p. 18-30. multiobjective clustering. in Proc. of IEEE Congress on Evolutionary
[20] Saha, S., and Bandyopadhyay, S. A multiobjective simulated annealing Computation (CEC 2005). 2005a: IEEE.
based fuzzy-clustering technique with symmetry for pixel classification [39] Hore, P., Hall, L., and Goldgof, D. A cluster ensemble framework for
in remote sensing imagery. in Proc. of 19th International Conference on large data sets. in Proc. of IEEE International Conference on Systems,
Pattern Recognition. 2008a. Man and Cybernetics. 2007.
[21] Saha, S., and Bandyopadhyay, S. Unsupervised pixel classification in [40] Hore, P., Hall, L.O., and Goldgof, D.B., A scalable framework for
satellite imagery using a new multiobjective symmetry based clustering cluster ensembles. Pattern Recognition, 2009.
approach. in Proc. of IEEE Region 10 Annual International Conference. [41] Fliege, J., Approximation Techniques for the Set of Efficient Points.
2008b. 2001, Universitt Dortmund, Germany.
[22] Xie, X.L., and Beni, G., A validity measure for fuzzy clustering. IEEE [42] Freitas, A.A., A Critical Review of Multi-Objective Optimization in
Trans. on Pattern Analysis and Machine Intelligence, 1991. 13 p. 841 Data Mining: a position paper. ACM SIGKDD Explorations
847. Newsletter, 2004. 6(2): p. 77-86.
[23] Jain, A.K., and Dubes, R. C., Algorithms for Clustering Data. [43] Brockhoff, D., and Zitzler, E., Are All Objectives Necessary? On
Englewood Cliffs. 1988: NJ: Prentice-Hall. Dimensionality Reduction in Evolutionary Multiobjective
[24] Pakhiraa, M.K., Bandyopadhyay, S., and Maulik U., Validity index for Optimization, in Parallel Problem Solving from Nature - PPSN IX,
crisp and fuzzy clusters. Pattern Recognition, 2004. 37 p. 487 501. Lecture Notes in Computer Science 2006, Springer.
[25] Handl, J., and Knowles, J., Multi-Objective Clustering and Cluster [44] Jaimes, A.L., Coello Coello, C.A. and Chakraborty, D. Objective
Validation, in Studies in Computational Intelligence 2006c, Springer. reduction using a feature selection technique. in Proc. of 10th Annual
[26] Mukhopadhyay, A., Bandyopadhyay, S., and Maulik, U. Clustering Conference on Genetic and evolutionary computation. 2008.
using multi-objective genetic algorithm and its application to image [45] Hughes, E.J. Evolutionary Many-Objective Optimisation: Many Once or
segmentation. in Proc. of IEEE International Conference on Systems, One Many? in Proc. of IEEE Congress on Evolutionary Computation
Man and Cybernetics. 2007. (CEC2005). 2005 Edinburgh, Scotland: IEEE.
[27] Handl, J., and Knowles, J. On semi-supervised clustering via [46] Brockhoff, D.a.Z., E., Objective Reduction in Evolutionary
multiobjective optimization. in Proc. of 8th annual conference on Multiobjective Optimization: Theory and Applications. Evolutionary
Genetic and evolutionary computation (GECCO2006). 2006a: ACM. Computation, 2009. 17(2): p. 135166.
[28] Handl, J., and Knowles, J., An evolutionary approach to multiobjective [47] Xu, R., and Wunsch, D., Survey of clustering algorithms. IEEE Trans.
clustering. IEEE Trans. on Evolutionary Computation, 2007. 11(1): p. on Neural Networks, 2005. 16: p. 645-678.
5676. [48] Zhu, D., A hybrid approach for efficient ensembles. Decision Support
[29] Matake, N., Hiroyasu, T., Miki, M., and Senda, T. Multiobjective Systems, 2009, in press.
clustering with automatic k-determination for large-scale data. in Proc. [49] Bandyopadhyay, S., Pal, S.K. and B. Aruna, Multiobjective GAs,
of Genetic and Evolutionary Computation Conference. 2007. Quantitative Indices, and Pattern Classification. IEEE Trans. on
[30] Santos, D.S., Oliveira, D.d. and Bazzan A.L.C., A Multiagent, Systems, Man and Cybernetics-Part B: Cybernetics, 2004. 34(5): p.
Multiobjective Clustering Algorithm, in Data Mining and Multi-agent 2088-2099.
Integration. 2009, Springer. [50] Miettinen, K., Introduction to Multiobjective optimization:
[31] Mukhopadhyay, A., Bandyopadhyay, S., and Maulik, U. Combining Noninterative approaches, in Multiobjective Optimization: Interactive
Multiobjective Fuzzy Clustering and Probabilistic ANN Classifier for and Evolutionary Approaches, J. Branke, Deb, K., and Miettinen, K.,
Unsupervised Pattern Classification: Application to Satellite Image Editor. 2008, Lecture Notes in Computer Science, Springer.
Segmentation. in Proc. of Congress on Evolutionary Computation. [51] Bong, C.W., and Wang, Y.C., A multi-objective hybrid metaheuristic
2008: IEEE. for zone definition procedure. International Journal of Services
[32] Qian, X., Zhang, X., Jiao, L. and Ma, W. Unsupervised texture image Operations and Informatics, 2006 1(1-2): p. 146 - 164.
segmentation using multiobjective evolutionary clustering ensemble [52] Miettinen, K., Ruiz, F., and Wierzbicki A. P., Introduction to
algorithm. in Proc. of Congress on Evolutionary Computation. 2008. Multiobjective Optimization: Interactive Approaches, in Multiobjective
[33] Faceli, K., De-Souto, M.C.P. and De-Carvalho, A.C.P.L.F. A strategy Optimization: Interactive and Evolutionary Approaches, J. Branke, Deb,
for the selection of solutions of the pareto front approximation in multi- K., and Miettinen, K., Editor. 2008, Lecture Notes in Computer Science,
Springer.

2010 IEEE Conference on Cybernetics and Intelligent Systems 155

Vous aimerez peut-être aussi