An Unsupervised Nonparametric and Cooperative Approach

for Classification of Multicomponent Image Contents

Akar Taher, Kacem Chehdi and Claude Cariou

University of Rennes 1, TSI2M, IETR/ENSSAT - 6, rue de Kerampont, 22300 Lannion, France

Keywords: Classification, Cooperative, Unsupervised, Nonparametric, Genetic Algorithm, Multicomponent, Image.

Abstract: In this paper an unsupervised nonparametric cooperative and adaptive approach for multicomponent image

partitioning is presented. In this approach the images are partitioned component by component and

intermediate classification results are evaluated and fused, to get the final partitioning result. Two

unsupervised classification methods are used in parallel cooperation to partition each component of the

image. The originality of the approach relies i) on its local adaptation to the type of regions in an image

(textured, non-textured), ii) on the automatic estimation of the number of classes and iii) on the introduction

of several levels of evaluation and validation of intermediate partitioning results before obtaining the final

classification result. For the management of similar or conflicting results issued from the two classification

methods, we gradually introduced various assessment steps that exploit the information of each component

and its adjacent components, and finally the information of all the components. In our approach, the

detected region types are treated separately from the feature extraction step, to the final classification

results. The efficiency of our approach is shown on two real applications using a hyperspectral image for the

identification of invasive and non invasive vegetation and a multispectral image for pine trees detection.

1 INTRODUCTION

The growth and the availability of multicomponent

images (e.g. hyperspectral images with hundreds of

spectral bands) which contain rich information have

opened new possibilities of applications in many

domains. In order to interpret the richness of this

information, a large diversity of image classification

approaches can be found in the literature. In

(Kermad and Chehdi, 2002), these approaches are

classified into two groups: non-cooperative and

cooperative approaches. Non-cooperative

approaches use a single classification method. These

methods can be supervised or unsupervised,

parametric or nonparametric. Supervised parametric

methods like Maximum Likelihood (ML) (Cox and

Snell, 1968), Support Vector Machines (SVM)

(Vapnik, 1998), Expectation Maximization (EM)

(Dempster et al., 1977), are the most commonly

used. These methods need a priori information to

accomplish the classification task. However the

required information is not available in all

application cases. The most frequently used

unsupervised nonparametric methods include: K-

Means (McQueen, 1967), LBG (Linde et al., 1980),

Fuzzy C-means (FCM) (Bezdek, 1981), and the

more recent Affinity Propagation (AP) (Frey and

Dueck, 2007). These methods do not require or

require little information a priori to accomplish the

classification task.

Another difficulty to the general problem of

segmentation or classification of multicomponent

images is the fact that applying several methods to a

single image never gives identical results. This

situation makes the quality assessment for the choice

of a particular method very difficult. Therefore,

cooperation between methods is highly desirable to

reliably partition images. The cooperation can be

done using three different schemes (Kermad and

Chehdi, 2002): sequential, parallel, or hybrid

cooperation. The drawback of the first scheme is the

sequential order of the applied methods which can

highly affect the results. Parallel cooperation

schemes use different classification methods at the

same time. A fusion stage introducing a validation

criterion is therefore required at the end of the

process. The main difficulty of this scheme is the

fusion process which requires a robust decision rule

to give optimal results. However, the application of

different classification methods in a parallel manner

263

Taher A., Chehdi K. and Cariou C..

An Unsupervised Nonparametric and Cooperative Approach for Classiﬁcation of Multicomponent Image Contents.

DOI: 10.5220/0004828502630270

In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 263-270

ISBN: 978-989-758-018-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

allows reducing the computing time. The third

cooperation scheme is the hybrid cooperation which

combines the two previous schemes. This third

scheme has essentially the same drawbacks as the

first one.

In the literature, many cooperative approaches of

multicomponent image partitioning are available.

Part of them uses supervised parametric methods

(Tarabalka et al., 2009), (Kalluri et al., 2010),

(Benediktsson and Kanellopoulos, 1999) but some

others use unsupervised methods (Forestier et al,

2010).

(Tarabalka et al., 2009) propose a parallel

cooperative approach that uses three classification

methods: Support Vector Machines (SVM),

Maximum Likelihood (ML), and ISODATA. A

sequential approach, presented by (Benediktsson and

Kanellopoulos, 1999), involves the cooperation of

Neural Networks and ML methods. Another recent

parallel approach proposed by (Kalluri et al., 2010)

uses only ML classification method to partition an

image. The partitioning process is repeated several

times, changing the features extracted from the

image each time, and then the results are fused to get

the final result. The drawback of these cooperative

approaches resides in the use of supervised and/or

parametric methods, which require the availability of

some prior information that is not available in all

applications cases. The approach proposed by

(Forestier et al, 2010) makes cooperation between

different unsupervised methods in parallel but

requires some background knowledge about the data

while fusing the results of the different methods.

The main ideas of the approach presented in this

paper are based firstly, on the use of unsupervised

nonparametric classification methods and, secondly,

on the management of conflicting partitioning

results. The developed approach belongs to the

family of parallel cooperation scheme.

This paper is organized as follows: the second

section describes the proposed approach, the third

section presents applicative experiments on different

real images and finally, the last section gives the

conclusions and provides some perspectives.

2 APPROACH DESCRIPTION

The proposed partitioning approach is composed of

four steps (for the first three steps, see Figure

1):

The First Step is the adaptive feature extraction,

in which the image is divided into two types of

regions, i.e. textured and non-textured. The adaptive

characterization of pixels, taking into account the

textured or non-textured nature of the region to

which they belong, is an essential step before the

classification. Indeed, the features dedicated to the

description of regions with low variance do not have

sufficient discriminating power for textured regions,

and vice versa.

The Second Step is the unsupervised parallel

classification, in which the image is partitioned

using two different unsupervised nonparametric

classification methods (FCM and LBG) where the

number of classes is estimated automatically. In this

step, the pixels belonging to textured or non-textured

regions are classified separately and in parallel using

appropriate feature sets.

The Third Step includes the results management

of the same component (monocomponent image)

which is done at two levels, firstly by validating

pixels which are coherently classified by the two

methods (FCM and LBG), and secondly by

processing conflicting classification results using a

genetic algorithm (GA). The objective function of

the genetic algorithm depends on between-class and

within-class disparities to evaluate and manage the

conflicting pixels between the partitioning results.

The last block of fusion is the union between the

results of textured and not textured regions

In the case of multicomponent images the above

three steps are applied independently on each

component.

The Fourth Step is the identification of similar

pixels between the classification results of adjacent

components. In this step the results from the

different components are grouped into subsets,

which are formed depending on the number of pixels

that are classified to the same class in different

components. Then these subsets are processed

independently to get one classification result for

each of them. The same process as in the third step

is used to get the final result of the multicomponent

image.

In the following subsections the approach is

described in details.

2.1 Adaptive Feature Extraction

This step is composed of two processes. In the first

one, the image is globally analyzed, in order to be

divided into two types of regions: textured and non-

textured. In the second process, features are

extracted taking into account the type of detected

regions.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

264

Figure 1: The general layout of the proposed approach (case of monocomponent image).

2.1.1 Region Nature Detection

The detection of regions’ types is done by local

extraction of the uniformity feature issued from the

co-occurrence matrix (Haralick, 1979). The

uniformity features are calculated using the

following equation (Rosenberger and Chehdi, 2003):

),(

, ii

ggP

Unif









(1)

Where L is the number of gray levels in the

monocomponent image, g

is an gray level, M is the

number of orientations θ

used to compute the co-

occurrence matrices, and P

d, θ

(.,.) are the entries of

the co-occurrence matrix obtained with inter-pixel

distance d.

The co-occurrence matrix is very time

consuming for an image in its original gray levels. In

order to reduce the gray levels number of an image

and while preserving the significant information at

the same time, we have used the multi-thresholding

method described in (Kermad et al., 1995).

In order to consider different texture scales, this

feature (uniformity) is extracted for each pixel using

five different sizes of an analysis window. These

extracted features are injected into an FCM classifier

and pixels are partitioned into two classes; by this

procedure, the textured and non-textured regions are

identified. This process enables the adaptativity of

feature extraction in the further processing steps.

To validate this procedure of region nature

detection we have applied it on a set of 30 images.

The database of images used includes images which

are composed of two types of regions: two textured

regions (textures are taken from the Brodatz album)

(Brodatz, 1966) and three non-textured regions.

Figure 2 shows a sample image and the

corresponding region nature detection map. The

Average Correct Detection Rate (ACDR) for all the

tested images is 98.60%.

(a)

(b)

(c)

(d)

Figure 2: Region nature detection results. (a) Image

ground truth mask (b) region type mask: white is textured,

black is non-textured, (c) Original sample image,

(d) Detected region types.

2.1.2 Feature Extraction

Two types of features are extracted to characterize

pixels in an image: the local average of pixel values

is a sufficient feature in the case where pixels belong

to a non-textured region; however in the case of a

textured region, several texture descriptive features

are needed. In (Rosenberger and Chehdi, 2003), 23

features are analyzed and 15 of them are selected

after removing the most redundant ones. These

features are:

 Mean, variance, skewness, and kurtosis.

 9 features issued from the co-occurrence

matrix (Haralick, 1979): contrast,

correlation, inverse difference moment, sum

average, sum entropy, entropy, first and

second information measure of correlation,

contour information.

Adaptive

feature

extraction

Textured regions

features

Non-textured regions

features

Class number

estimation

and unsupervised

classification using

(FCM, LBG)

Class number

estimation

and unsupervised

classification using

(

FCM

LBG

)

Evaluation and

conflict

management

Classification

result

Fusing

results

Evaluation and

conflict

management

Mono

component

image

AnUnsupervisedNonparametricandCooperativeApproachforClassificationofMulticomponentImageContents

265

 2 features from the curvilinear integral.

All the features are calculated locally by sliding

an analysis window with maximum overlapping of

size WxW pixels (W odd). We have chosen a

window size of 3x3 pixels to compute the local

average for the non-textured regions, and extended it

to 9x9 pixels for features of the textured regions.

Identifying pixels belonging to one among two

types of regions brings two advantages, i.e. a natural

adaptation of feature extraction, and a computation

time gain in feature calculation.

2.2 Unsupervised Parallel

Classification

In this step the pixels in the textured and non-

textured regions are processed independently and in

parallel. The features extracted are injected into two

classification methods (FCM and LBG) chosen to

cooperate with each other due to their respective

fuzzy and hard decisions rules. These two methods

are used in many applicative domains, and have

shown their efficiency (Havens et al., 2012),

(Huang

and Xie, 2010). Their common advantage is that

they are both nonparametric, but both of them

require the prior knowledge of the number of classes

which is not known in most practical cases.

As our approach is considered to be completely

unsupervised, a step of class number estimation is

introduced. The pixels are classified in an iterative

manner using FCM. More precisely, the pixels are

classified using FCM several times by increasing the

number of class k, starting from k=2. The result

obtained at the end of each iteration is evaluated

using combined within-class and between-class

disparities as follows:

)()(1

))(),((

IDID

IDIDF





(2)

The within-class disparity

)(ID

quantifies the

homogeneity of each class obtained in the

partitioning result. Similarly, the global between-

class disparity

)(ID

measures the disparity between

the classes (Rosenberger and Chehdi, 2003). This

criterion provides some advantages: firstly it is

unsupervised; and secondly it adapts itself

automatically to the nature of the regions (textured,

non-textured). Finally it is composed of a

combination of two measures that both control

efficiently the issue of under and over classification.

The optimum value of the number of classes is

the one which maximizes criterion (2). Then the

estimated class number by FCM classifier is directly

used in the LBG method.

The motivation for applying FCM before LBG is

that, it is more robust in the estimation of the

number of classes. Actually, the LBG method gives

under-partitioning as shown in Figure 4.

This process is tested on the set of images

described previously in section 2.1.1. The average

correct class number estimation over the tested

image set is 90%. This rate is coherent because in

some cases there are high fluctuations within a class

so that it is detected as more than one class. For

example, the class labelled as “1” in the image in

Figure 3-(a) is composed of wood, where a part of

this class is defected; it is clear from visual

inspection that the area inside the highlighting red

oval (wood defect) is not the same as the rest of the

class. This class is detected as two classes by our

method, which is actually true.

We can point out that if this kind of information

were accounted for in the ground truth of the image

set, the correct class estimation rate would be greater

than 90%.

Figure 3: Example of estimation the number of classes.

(a) Original image, (b) Classification result (6 classes).

2.3 Evaluation and Management of

Conflicts

In the partitioning process each classification

method generates a different partition from the same

data. Therefore, effective evaluation criteria are

important to provide the end users a degree of

confidence in clustering results. These assessments

should be objective and have no preference to any

algorithm (Xu and Wunsch, 2005).

2.3.1 Monocomponent Image Case

To validate the classification results obtained from

FCM and LBG, a two-level evaluation process is

applied. First, the pixels that are classified to the

same classes by both methods are considered

directly as valid pixels, and reported to the final

partitioning result. These pixels do not enter the

second level of evaluation, which dramatically

considerably reduces the complexity of processing.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

266

The pixels that are classified to different classes by

the two methods are considered as invalidated, and

are subject to a second evaluation process using an

objective function optimized by a GA. The objective

function used is based on within-class and between-

class disparities (see Equation (2)).

In the GA (Holland, 1992), each pixel is

considered as a gene and each classification result is

considered as a chromosome. In our case, each

chromosome is composed of the invalidated pixels

only. Since there are two classification methods, the

initial population is composed of two chromosomes.

This population is evolved using genetic operations,

where only selection and cross-over operations are

considered. The genetic algorithm stops when no

better chromosomes are created, or the quality

difference between the best chromosomes of the two

last generations is smaller than ɛ, which guarantees

the termination of the GA. In our experiments the

value of ɛ is set to 10

-10

. The selection operation

used in this approach is the fitness proportionate

selection (Neumann et al. 2009). With this selection

operation type there is a chance for some weaker

solutions to survive the selection process; this is an

advantage, as though a solution may be weak, it may

include some information which could prove useful

following the recombination process. The type of

cross-over operation used in the approach is the

uniform cross-over (Syswerda, 1989), which uses a

fixed mixing ratio between two parent

chromosomes. The advantage of this cross-over

operation type is that it enables the parent

chromosomes to exchange at the gene level rather

than at the segment level. At termination of the GA,

the best evaluated chromosome in the population is

considered as the final result for the conflicting

pixels. Eventually, these pixels are grouped with the

valid pixels from the first level, to get the final

partitioning result for a single component).

To have the complete classification result of the

monocomponent image, the fusion in the last block

of Figure 1, is a union between the partitioning

results of textured and no textured regions.

To prove the validity of this process, we have

assessed it on the database of images previously

described in section 2.1.1. Figure 4 shows a sample

image and the obtained results including the detected

region natures, the individual results of LBG and

FCM, the result of our cooperative approach, and the

corresponding average correct classification rates

(ACCR). The number of classes in this experiment is

correctly estimated to 5 classes. In this example, the

result given by the LBG method only mixes up two

regions of the image, yielding a low correct

classification rate, while the result obtained by FCM

is clearly more robust. Another remark is that some

pixels from the LBG method result are classified in

the correct class, which is not true with the FCM

method. The application of the cooperative approach

has kept the pixels correctly classified and reassigns

those which were not previously correctly classified.

(a)

(b)

(c)

(d)

(e)

(f)

Figure 4: Results of mono-component image. (a) Original

image, (b) Region nature detection (white: textured, black:

non-textured), (c) LBG result (ACCR: 78.20%), (d) FCM

result (ACCR: 98.58%), (e) Validation map (white: valid,

black: invalid), (f) Our approach result (ACCR: 98.63%).

The global average correct classification rate, for the

set of 30 test images is: 94.71% for FCM method,

88.31% for LBG method and 97.19% for our

approach.

2.3.2 Multicomponent Image Case

In this case the results from the different

components are fused to get the final classification

result.

The steps described in sections (2.1, 2.2 and

2.3.1) are applied to each component in the image

independently to obtain N partition results, where N

is the number of components of the image. To get

the final partition result, the results of different

AnUnsupervisedNonparametricandCooperativeApproachforClassificationofMulticomponentImageContents

267

components are compared, and adjacent components

with similar classification results are grouped within

the same subset. The contents of each subset are

fused independently. At the beginning the first

component result is taken as reference and compared

with the adjacent components result. The reference

component is changed if the number of identical

pixels decreases. For example, if the first component

result is compared with the second component result

and some percentage of the pixels where found to be

classified identically, then the first component result

is compared with the third component result, if the

percentage of the identical classified pixels are

greater than this percentage, the reference

component remains unchanged and compared with a

further component result; if not, the first and second

component results are considered as one subset and

the third component result becomes the reference

component, and the same procedure is repeated until

the last component is processed.

The component results in each subset are fused

separately, and then the results of the subsets are

fused to get the final partitioning result of the

multicomponent image. GA is used in the fusion

process where the objective function is the same as

the one described previously.

To validate this approach of evaluation and

fusion, it is applied on a hyperspectral test image

which is constructed from regions of interest of a

real image. The ground truth of the image is

available (used for estimation of the results only).

The original hyperspectral image was collected by

an AISA Eagle sensor on October 1

2010 over the

region of Cieza in southeastern Spain. It was

acquired at 0.5 meter spatial resolution in 62 spectral

bands within the range [400, 970] nm. The data used

to construct the test image are taken randomly from

5 different regions of the original image. The 5 land

covers are: Water, Pinus halepensis, Peach trees,

Arundo donax, and Buildings. The estimated number

of class for this test is 5 classes. The result for this

test is shown in Figure 5.

3 ASSESSMENT ON REAL

APPLICATIONS

Our classification approach was also evaluated on

two real applications using respectively a

hyperspectral image for identification of invasive

and non invasive vegetation in the region of Cieza

(Spain) and a multispectral image for pine trees

detection in the region of Baabdat (Lebanon). For

(a)

(b)

(c)

(d)

(e)

Water

Pinus halepensis

Peach trees

Arundo donax

Buildings

Figure 5: Hyperspectral test image classification results.

(a) Original image (RGB components over 62 for

visualizing only), (b) ground truth masks, (c) FCM result

(ACCR: 91.68%), (d) LBG result (ACCR: 69.59%),

(e) Our approach result (ACCR: 97.95%).

these experiments, a ground truth data is available,

which allows us to assess the performances of our

approach.

The characteristics of the hyperspectral image

are described in section 2.3.2. The spatial size of this

image is1000 lines by 1000 columns.

The ground truth of the image includes six

different classes, namely: Phragmites australis,

Arundo donax, Tamarix, Ulmus minor, Pinus

halepensis, and Peach trees.

The multispectral image was acquired by the

Earth observation satellite Ikonos on July 11, 2005,

in the region of Baabdat (Lebanon) and it is used to

detect pine trees. The ground pixel size of this three

components (RGB) image is 0.8m.

To assess our unsupervised approach on these

two applications, the correct classification rates are

calculated using available ground truth areas.

For the detection of invasive and non invasive

vegetation, Figure 6 and Table 1 respectively show

the result of our classification approach and the

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

268

corresponding confusion matrix. The estimated

number of classes for this experiment is 6 classes.

The average correct classification rate for FCM,

LBG, and our approach are: 98.70%, 98.55%, and

99.13% respectively.

Concerning the detection of pine trees

application, the ground truth data contains 11736

pixels, with 11410 of them labeled as pine trees. In

this test, the estimated number of classes is 2. The

average correct classification rates (ACCR)

calculated only on the ground truth zones is: 96.91%

for FCM method, 97.09% for LBG method, and

lastly 97.22% for our cooperative approach.

A key point to mention is that, in all the

experiments, the ground truths provided with the

image data is only used for evaluating the results of

our approach, since the approach is unsupervised

and does not require any training; or any other

knowledge about the data.

4 CONCLUSIONS

This paper addresses the design of a cooperative and

adaptive classification approach for partitioning

mono or multicomponent images. The proposed

approach is non parametric and totally unsupervised.

The characterization of pixels according to the

nature of the region to which they belong, reduces

the computation burden of feature extraction, and

makes their processing independent in the

classification and result management steps.

The automatic estimation of class numbers

allows to better identify or classify the contents of

images. Indeed forcing a classifier to a subjective

fixed class number generally causes false or less

accurate classification results.

The systematic validation of the results from the

classification methods used, and the assessment of

only conflicting results before their validation leads

to consistent and reliable results. Several evaluations

on monocomponent and multicomponent (multi and

hyperspectral) images show the robustness of the

proposed approach results compared to those

obtained without cooperation.

The integration of the spectral information in the

feature extraction step remains as a perspective for

this work.

(a)

(b)

(c)

(d)

Phragmites australis Ulmus minor

Arundo donax Pinus halepensis

Tamarix Peach trees

Figure 6: Detection of invasive and non invasive

vegetation results from hyperspectral image. (a) Original

Image (3 components over 62 for visualizing only), (b)

Ground truth masks, (c) Pixels of ground truth to classify,

(d) Our approach classification result (ACCR: 99.13%)

ACKNOWLEDGEMENTS

This study was conducted in the framework of the

franco-lebanese cedre programme n° 2010/ef/l6.

Authors wish to thank the national Council for

Scientific Research, Centre for Remote Sensing

Lebanon, for providing the dataset on pine trees.

They also thank Infraestructura y Ecología S.A.-

Aretech Group for providing the dataset on invasive

and non invasive vegetation species.

REFERENCES

Benediktsson J. A., Kanellopoulos I., 1999. Classification

of multisource and hyperspectral data based on

decision fusion. IEEE Trans. on Geoscience and

Remote Sensing. 37(3), 1367-1377.

Bezdek J. C., 1981. Pattern recognition with fuzzy

objective function algorithms. Kluwer Academic

Publishers Norwell, MA, USA.

Brodatz P., 1966. Textures: A Photographic Album for

Artists and Designers, Dover Publications, New York.

AnUnsupervisedNonparametricandCooperativeApproachforClassificationofMulticomponentImageContents

269

Table 1: Confusion Matrix of classification result using the proposed approach for detection of invasive and non invasive

vegetation: correct classification rate in %, (.): number of pixels. Average correct classification rate: 99.13 %.

Ground truth classes (number of pixels)

Classes predicted

by our approach

Phragmites

australis

(544)

Arundo donax

(4200)

Tamarix

(162)

Ulmus minor

(764)

Pinus

halepensis

(274)

Peach trees

(3040)

Phragmites australis 99.08% (539) 0.16% (7) 2.47% (4) 0.52% (4) 0.73% (2) 0.33% (10)

Arundo donax 0 99.84% (4193) 0 0 0 0

Tamarix 0 0 97.53% (158) 0 0 0

Ulmus minor 0,55% (3) 0 0 99.48% (760) 0 0.07% (2)

Pinus halepensis 0.37% (2) 0 0 0 99.27% (272) 0

Peach trees 0 0 0 0 0 99.6% (3103)

Cox D. R. and Snell E. J., 1968. A general definition of

residuals. Journal of the Royal Statistical Society,

Series B. 30(2), 248-275.

Dempster A. P., N. M. Laird and D. B. Rubin, 1977.

Maximum likelihood from incomplete data via the EM

algorithm. Journal of the Royal Statistical Society,

Series B. 39(1), 1-38.

Forestier G., Gancarski P., and Wemmert C., 2010.

Collaborative clustering with background knowledge.

Data and Knowledge Engineering. 69(2), 211-228.

Frey B. J. and Dueck D., 2007. Clustering by passing

messages between data points. University of Toronto

Science 315, 972–976.

Haralick R., 1979. Statistical and structural approaches to

texture. Proceedings of the IEEE. 69, 786-804.

Havens T. C., Bezdek J. C., Leckie Ch., Lawrence O., and

Palaniswami M., 2012. Fuzzy C-means algorithms for

very large data. IEEE Trans. on Fuzzy Systems. 20(6),

1130-1146.

Holland J. H., 1992. Adaptation in natural and artificial

system. 2nd Edition. Cambridge, MA, USA: MIT

press.

Huang B., Xie L., 2010. An improved LBG algorithm for

image vector quantization. 3rd IEEE International

Conference on Computer Science and Information

Technology (ICCSIT).

Kalluri H. R., Prasad S., and Bruce L. M., 2010. Decision-

level fusion of spectral reflectance and derivative

information for robust hyperspectral land cover

classification. IEEE Trans. on Geoscience and Remote

Sensing. 48(11), 4047-4057.

Kermad Ch., Chehdi K., 2002. Research of an automatic

and unsupervised system of segmentation. Image and

Vision Computing Edition Elsevier. 541-555.

Kermad Ch, Chehdi K, Cariou C, 1995. Segmentation

d’images par multi-seuillage et fusion des régions

labellisée minimisant un critère de similarité. In

Quinzième Colloque GRETSI, 2, 641-644. Juan-Les-

Pins, France.

Linde, Y., Buzo, A., Gray R. M., 1980. An Algorithm for

vector quantizer design. IEEE Transactions on

Communications. 28, 84–94.

McQueen J., 1967. Some methods for classification and

analysis of multivariate observations. In Proceedings

of the Fifth Berkeley Symposium on Mathematical

Statistics and Probability. 1, 281-297.

Neumann F., Oliveto P. S., Witt C., 2009. Heoretical

analysis of fitness-proportional selection: landscapes

and efficiency. In Proc. of 11

(GECCO), 835-842.

Rosenberger Ch., and Chehdi K., 2003. Unsupervised

segmentation of multi-spectral images. In Proc.

International Conference on Advanced Concepts for

Intelligent Vision Systems, Ghent, Belgium.

Syswerda G., 1989. Uniform crossover in genetic

algorithms. In Proc. 3rd International Conference on

Genetic Algorithms, 2-9.

Tarabalka Y., Benediktsson J. A., and Chanussot J., 2009.

Spectral–spatial classification of hyperspectral

imagery based on partitional clustering techniques.

IEEE Trans. on Geoscience and Remote Sensing.

47(8), 2973-2987.

Vapnik V., 1998. A tutorial on support vector machines

for pattern recognition. Statistical Learning Theory,

New York: Wiley, 1998. C. Burges, Data Mining and

Knowledge Discovery. 2, 121-167.

Xu R., Wunsch II D., 2005. Survey of clustering

algorithms. IEEE Trans. on Neural Networks. 16(3),

645-678.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

270