SETIT 2005

3rd International Conference: Sciences of Electronic,

Technologies of Information and Telecommunications
March 27-31, 2005 – TUNISIA

Ant Colonies For MRF-Based Image Segmentation

Salima Ouadfel1,2 ,Mohamed Batouche2 and Said Talhi1
1
Computer Science Department, University of Batna.
[email protected]
[email protected]
2
Computer Vision Group, LIRE Laboratory, University of Constantine
[email protected]

Résumé: This paper presents HACSEG, a new ant algorithm for the image segmentation based on the Markov Random
Field (MRF) and a modified version of the Ant Colony System algorithm coupled with a local search. HACSEG
algorithm differs from other ant algorithms proposed for image segmentation, in the way that each artificial ant is
associated with a particular partition that is modified using pheromone trails and heuristic information unlike to build a
completely new partition using a iterative constructive process. The new partitions found by ants are then optimized
using a local search algorithm. Pheromone trails are updated according to the quality of the partitions found by the best
ant. A diversification phase is also used to diversify the search. The experimental results presented outperforms those
obtained with other methods.
.
Mots clés: Image segmentation, Clustering, Markov Random Field, Ant Colony System., local search.

many local minima in the solution space.

1 Introduction As a result, techniques such as iterated conditional
Image segmentation is a low-level image method (ICM) (Besag, 1986), simulated annealing
processing task in a vision system. It has been the (SA) (Geman & al, 1984), (Hu & al, 1992), (Gunnels
subject of intensive research, and a wide variety of & al, 1994) and genetic algorithm (GA) (Andrey & al,
image segmentation techniques have been reported in 1998), (Kim & al, 1998), Huang & al, 1995) has been
the literature. A good review of these methods can be often used as solution for this computational
found in (Pal & al, 1993). Among them, Markov complexity.
random field (MRF) is one of the most frequently Ant Colony Optimization (ACO) metaheuristic
utilized techniques (Andrey & al, 1998), (Chellappa & (Dorigo & al, 1999), (Maniezzo & al, 2001) is a multi-
al, 1993), (Chen & al, 1995), (Dubes & al, 1990), agent metaheuristic for combinatorial optimization
(Kato, 1994), (Kervrann & al, 1995), (Panjwani & al, and other problems. It is inspired by the capability of
1995). MRF has been shown to be quite successful for real ants to find the shortest path between their nest
image segmentation because of its ability to and a food source. In ACO, a colony of artificial ants
characterize spatial relations among image pixels by build new solutions of the problem within a stochastic
conditional probability over small neighborhoods of iterative process, by adding solution components to
pixels. The image is segmented by maximizing the a partial solutions using a combination of heuristic
posteriori probability (MAP) of the labeling space information and an artificial pheromone trail. The
given the image data (Li, 1995). Within this pheromone trail is reinforced according to the quality
framework, the segmentation process is expressed as of the solutions built by the ants.
the problem of finding the optimum value of an
energy function (Lin & al, 1993), (Panjwani & al, In previous works (Ouadfel & al, 2003a), (Ouadfel
1995). This is combinatorial optimization problem & al, 2003b), we have proposed ACS-MRF, a MRF
because of the large search space. Moreover, the model based image segmentation using a hybrid ant
energy function is usually non-convex and exhibits colony system algorithm. In this paper, we investigate
the capability of HACSEG, a new ant algorithm for
SETIT2005

image segmentation, that differ in (1) manipulation of σ xs : the deviation value of the cluster xs
solutions, (2) the way pheromone is exploited and
laid, (3) the use of a heuristic information to guide Vc (x ) : the potential function for clique c
ants and (4) the introduction of a diversification phase C : the set of all cliques over the image.
to diversify the search in the solutions space.
Experimental results show that the HACSEG obtained
good quality segmentation results and outperforms A clique is a set of pixels that are neighboring of
ACS-MRF and others methods. one another. In this paper we consider only the pair-
site clique potentials of 8-neighborhood system, with
The paper is organized as follows. Section 2 the form V(x1, x2)= −β if x1 = x2 and 0 otherwise. β is
presents a brief review on image modeling using a positive parameter and the larger β, the larger is the
MRF. Section 3 describes ACO paradigm. Section 4 influence of the neighboring pixels.
presents HACSEG for MRF based image
segmentation. In section 5 we present the Minimization of the global energy function, is hard
experimental results and we compare HACSEG with optimization problem because the number of possible
other heuristics. Finally a conclusion is drawn in label configurations is generally very large and
section 6. moreover, the energy function may contains local
minima (Li, 1995).

2 Image segmentation using Markov

Random Field 3 Ant Colony Optimization
Ant Colony Optimization (ACO) is a population
The MRF was introduced in image analysis by
based approach attribute to Marco Dorigo in
Geman and Geman (Geman & al, 1984). MRF is a
collaboration with Alberto Colorni and Vittorio
stochastic process in which spatial relations within the
Maniezzo (Dorigo & al, 1991), (Bonabeau & al, 1999)
image are included in the labeling process through
and inspired by the foraging behavior of ant colonies
statistical dependence among neighboring pixels. Let
concerning in particular how they can find shortest
Y ={ys / ∀s ∈ S } designate an observation field
paths between food sources and their nest without
defined on a rectangular lattice S. Let the label
using visual cues. Ants foraging for food lay down
field X ={X s / ∀s ∈ S }defined on S and the set of
labels Λ ={0,...L −1} of the pixel s. Realization of
quantities of a volatile chemical substance named
pheromone, marking their path that it follows. It has
fields Y and X will be denoted by y ={ys / ∀s ∈ S} and
been observed that the more ants use a particular path,
x ={xs / ∀s ∈ S }. (X, Y) is a Markov random field on
the more pheromone is deposited on that path and the
S with respect to a neighboring system
more it becomes attractive and thereby reinforces it
N ={N s , s ∈ S }, where N s is the set of pixels
with a further quantity of pheromone. The process is
neighboring s.
) thus characterized by a positive feedback loop where
Our goal is to find the best estimated x for x given the probability that an ant chooses a path increases
y. According to the Maximum A Posteriori (MAP), with the number of ants choosing the path at previous
criterion x is obtained by minimizing the global times and with the strong of the pheromone
energy function U(y,x). concentration laid on it (Dorigo & al, 1997a).

) Consider the experimental setting shown in Figure.

x=arg min{U(y, x)} (1) 1. If an obstacle is suddenly placed on an established
path leading to a food source, ants will initially go
right or left in a seemingly random manner, but those
If we make the assumption that the image data are choosing the side that is in fact shorter will reach the
conditional independent and that Y is obtained by food more quickly and will make the return journey
adding an identical independently distributed (i.i.d.) more often. The pheromone on the shorter path will
Gaussian noise, and according to the Hammersley- therefore be more strongly reinforced and will
Clifford theorem (Besag, 1986) the energy U(y,x). is eventually become the preferred route for the stream
formulated as follows (Ouadfel & al, 2003b): of ants.

U(x) = ∑
(
 ys −µxs 2
+
) ∑log(σ )+∑V (x)
xs c

(2)
 s∈S 2σ xs
2
s∈S c∈C 

where

µ xs : the mean value of the cluster xs Figure.1. Ants facing an obstacle

SETIT2005

This indirect cooperation and stigmergetic HACSEG. HACSEG is a hybridization of a modified

communication has been used to solve difficult version of Ant Colony System (ACS) with a local
discrete combinatorial problems with a new paradigm search method.
Ant Colony Optimization (ACO). Figure. 2 presents
In HACSEG, each individual ant is associated with
the generic ACO algorithm. The fundamental
a particular partition and it uses pheromone trails and
approach underlying ACO is an iterative process in
heuristic information to perform modifications on
which a population of artificial ants collectively search
pixels’s labels, in the spirit of a neighborhood search,
for good quality solutions to discrete combinatorial
unlike (Ouadfel & al, 2003a) that uses pheromone trail
optimization problems. During each iteration, ants
information to construct a complete partition. During
repeatedly construct candidate solutions by adding
each iteration, the ant k will modify its partition x k by
components to a partial solution. Partial solutions are
performing partition swapping in the following way:
seen as the states and the ant moves from one state to
first a pixel s is chosen at random, then another pixel
another to a more complete partial solution according
s’ from its immediate neighbourhood N s is selected
to a probabilistic state transition rule (Maniezzo & al,
to be labelled with the same label as s, using a local
2001). The state transition rule depends on an artificial
heuristic information and the pheromone trails, The
pheromone trail τ representing experience gathered by
new partitions found by ants are then optimised using
ants in previous iterations and a heuristic information
a local search algorithm. Pheromone trails are then
η that represent a priori information of the given
updated according to the quality of the partitions
problem. Once all ants have built a solution,
found by the best ant. A diversification phase is also
pheromone trails are updated and the amount of
introduced when partitions seem not to be improving
pheromone deposited is a function of the quality of the
any more. It consists of a re-initialisation of both the
solution constructed. The goal of this update process is
pheromone trail matrix and the solutions associated to
the increasing the probability of choosing the moves
the ants (Gambardella & al, 1999).
that were part of good solutions, while decreasing all
others.
To apply ACO algorithm to our problem, the
following steps have to be taken: (1) The definition of
1- Initialization of the pheromone trails
the problem solution; (2) The choice of a suitable
2- For Imax iteration repeat
representation for artificial pheromone; (3) Assigning
a. For each Ant repeat
a heuristic preference to generated solutions at each
i. Solution construction using the
time step (i.e., select new components by the ants); (4)
pheromone trails and the heuristic information
Defining of the solution modification procedure.
ii. Update the pheromone trails
Problem solution representation
Figure. 2. A generic ACO algorithm
For the problem considered in this study, the solution
The first ant algorithm, called Ant System (AS) is given by the best partition x* , which corresponds
was proposed by Dorigo and applied to the traveling to a correct labeling of image pixels with respect to
the contextual constrains.
salesman problem (TSP) (Dorigo & al, 1991), (Dorigo
& al, 1997b). Since, the basic AS algorithm was
Artificial Pheromone trail
further improved concerning the transition rule, the
pheromone trail updating and the use of a local search, The artificial pheromone trail is numeric
leading to more elaborate variants ants algorithms that information encoded as a matrix of dimension (N, M)
were experimented on a broad range of hard where N is the number of image pixels and M=8 if we
combinatorial optimization problems including consider only the 8-neighborhood. Each element of
quadratic assignment (Maniezzo & al, 1994), (Taillard the matrix τ (s, j ) represents the amount of artificial
& al, 1997), (Gambardella & al, 1999), graph-coloring pheromone associated to the corresponding pair and
(Costa & al, 1997), (Bertelle & al, 2003), vehicle indicated the degree of desirability of setting the two
routing (Bullnheimer & al, 1997), telecommunication pixels in the same class in a solution, and made
networks (Di Caro & al, 1998) and sequential ordering available by previous attempts of other ants.
(Gambardella & al, 1997), solving the maximum clique
(Solnon & al, 2004), clustering (Trjeos & al, 2004), Heuristic function
image processing domain (Ramos & al, 2002), An important feature of an ACO implementation is
(Ouadfel & al, 2003a), (Ouadfel & al, 2003b), the choice of a good heuristic, which will be used in
(Meshoul & al, 2003). combination with the pheromone information to build
solutions. It guides the ants’ probabilistic solution
4 Ant Colonies For Image Segmentation construction/modification with problem specific
knowledge. In the case of this study, the heuristic
In this paper, the segmentation problem is function η should measure the degree of similarity
formalized as an optimization problem of the energy between two pixels. Therefore, the ant decides to label
function. For this, we use a new ant algorithm two pixels with the same label using the correlation
SETIT2005

between intensity patterns in a local neighborhood of information by moving pixels from one class to
the two pixels. The heuristic function values are another. N/2 moves are applied as follows. First a
computed from the linear correlation coefficient corr, pixel s is selected at random, next another pixel s’
once for all pixels in a preprocessing step and are from its immediate neighbors N s is selected to be
recorded in a matrix of dimension (N, M), where N is assigned to the same class as the pixel s, with two
the dimension of the image and M=8 if we consider different policies: with a probability q0, ant exploits
only the 8-neighborhood. The values of the heuristic the information contained in the pheromone trail and
function η are normalized to the range [0,1] and are heuristic information, while with probability (1-q0),
computed as follows: ant k explores the solution space.

corr(s,s')+1 (3) (3) More precisely, let q a random number uniformly

η(s,s')=
2 [ ]
distributed in 0,1 and q0 a fixed probability. If ant k
has selected pixel s with xsk = l , we have:
where corr(s,s') represents the linear correlation • If q ≤ q0 then, pixel s'∈ N s is chosen such as:
coefficient between a neighborhood around the pixel s
and a neighborhood around the pixel s’.
η(s,s') .τ k (s,s') is maximum
β γ
(4)
The different steps of HACSEG algorithm are
detailed bellow:
1- Generate K random partitions x k ; • Otherwise pixel s’ is chosen with the probability :
2- Associate each partition x k to one ant ;
3- Initialize the pheromone trail matrix;
η(s,s') .τ k (s,s')
β γ
4- Compute the heuristic function values; (5)
P(s,s')=
5- For max_iterations repeat
a. For each ant k =1,....., K do ∑η(s, j ) .τ (s, j )
β k γ

j∈N s
i. Modify the partition x k ;
ii. Apply a local search to the where
modified partition; o η(s,s') is the local heuristic between the
b. Record the global best solution two pixels s and s’ .
x gb =min (U(x k )) ; o τ k (s,s') is the desirability to assign the
k =1,...., K
pixels s and s’ to the same class in the
c. Update the pheromone trails; new partition of ant k.
6- If a number of iterations applied without
amelioration of x gb then apply Parameter q0 allows to modulate the degree of
Diversification. exploration (q0 ≈ 0) versus exploitation (q0 ≈ 1) and
to choose whether to concentrate the search of the
Figure. 3. HACSEG algorithm for image segmentation algorithm on the best solutions (greedy search) or to
explore the search space.
In the following, we present into details our
Pheromone trail update
implementation of HACSEG algorithm, step by step.
In HACSEG, as in (Ouadfel & al, 2003a) the
Initialization of partitions update of the pheromone trails is done only by the best
ant with the best quality found since the beginning of
Initially, we associate each ant k with a randomly
the algorithm. First, all the values contained in the
chosen partition x k such that each pixel s is given a
pheromone trail are reduced using the evaporation
random label from the set Λ . Theses partitions are
process (as in the real ants) according to the rule.
then optimized using ICM algorithm.

Pheromone trail initialization

From the K partitions generated in the previous τ(s, s')=(1− ρ).τ(s, s') (6)
phase, we record the global best partition
x gb =min (U(xk )) . For each pair (s, s’), s∈ ' N s ,we
where 0 < ρ < 1 is the evaporation parameter
have k=1,....,
chosenK to set τ(s, s') := τ 0 and
which has an influence on the exploratory behaviour
τ 0 = 1 (N.U(x gb)) , where N is the number of image
of ants. A value of ρ close to 0 implies that the
pixels.
influence of the pheromone will be efficient a long
Modification of the partitions time, while a value of ρ close to 1 implies a high
degree of evaporation and so the algorithm takes
During each iteration, the ants will modify their longer to find good solutions
partitions using the pheromone trails and heuristic
SETIT2005

In the second phase, only the globally best ant, the partitions.
one that constructed the clustering with the minimum
energy function defined in Eq (2) from the beginning 5 Experimental Results
of the algorithm, is allowed to update the pheromone
trails using the following update formula HACSEG is compared with the previous algorithm
presented in [Ouadfel & al, 2003a] which is also based
on ant colonies, and with other methods based on
τ(s, s')=τ(s, s') + ρ.∆τ(s, s') (7) genetic algorithm and simulated annealing. For the
comparison, we use cerebral magnetic resonance
(MR) images with differents levels of noise and
inhomogeneities (Figure. 3.) and real images
where
considered in (Ouadfel & al, 2003A).

 1 MR images
if (s, s') ∈ x gb (8)
U(x gb) Brain matter can generally be categorized as white
∆τ(s, s') = 
0 matter (WM), gray matter (GM) and cerebrospinal
 otherwise fluid (CSF). Tests have been done on the
computationally synthesized Brainweb phantoms
available on the site Brainweb:
Local search http://www.bic.mni.mcgill.ca/brainweb/,containing 0,
Local search has shown to be effective to solve 3 and 5% of noise and have an inhomogeneity of 20,
large optimization problems. Adding local search to or 40%, to compare the following segmentation
ant algorithms yields a faster convergence of the algorithms : (1) SA, (2) GA, (3)ACS-MRF, (4)
algorithm and an earlier detection of high quality HACSEG
solutions. The basic idea is to construct a complete
labeling and then iteratively refine it by applying a Inhomogeneity Inhomogeneit
neighborhood examination with fist improving =20% y =40%
strategy.
For each ant solution, we use a simple iterative
improvement local search that starting from an initial Noise=0%
solution iteratively generates a better solution from its
neighborhood. The neighborhood for a current
clustering x is the set of candidate clustering N(x)
that can be reached from x by making small
modifications on the labels of pixels. Iterative
improvement has been implemented using the first
improvement, i.e., the first improving move found is Noise=3%
accepted. For each solution x, we evaluate the
neighboring candidates solutions and the first
x' ⊂ N(x) for which U(x') < U(x) is selected. The
neighborhood of a solution is defined by a set of
solutions that can be obtained by exchanging the
labels of two pixels (Franti & al, 2000).
Noise=5%
Diversification
If x gb , the best global partition has not been
improved during a number of iterations which
indicates that the search is stuck or has prematurely
Figure. 3..Brain phantoms with different values of noise
converged the diversification step is started in order to
and inhomogeneities.
diversify the search. The pheromone trails are
completely re-initialised: All the values of the The parameters of SA GA and ACS-MRF are directly
pheromone trail matrix are set to the same value. and taken form (Ouadfel & al, 2003a). The parameters of
new partitions are generated for k – 1 ants expect for HACSEG are set identical to ACS-MRF and are
the keme ant for which we associate the best global tabulated in table 1
partition x gb . A matrix Freq ,
( )
(Freq s,s'),s∈S,s'∈N s is used to store the number
of times each pair (s,s’) has been incorporated into the
ant solutions. During the diversification phase, we Algorithms Parameters Value
have used assignments that have not been incorporated T0 2
into ants’solutions frequently to generate new
SETIT2005

SA Tm 0.9 20% 40% 20% 40% 20% 40%

Ni 100 LCR 0.86 0.87 0.86 0.85 0.82 0.80
Nmax 3000 (1) GM 0.91 0.89 0.88 0.88 0.86 0.84

N 30 WM 0.90 0.86 0.88 0.84 0.83 0.81

GA Pc 0.8 LCR 0.85 0.89 0.87 0.86 0.82 0.81

Pm 0.01 (2) GM 0.96 0.90 0.90 0.88 0.87 0.85

WM 0.91 0.86 0.88 0.84 0.83 0.82

Nmax 1000
LCR 0.89 0.87 0.87 0.85 0.86 0.82
Q0 0.6
ACS-MRF α 0.01 (3) GM 0.94 0.89 0.90 0.89 0.88 0.86

WM 0.93 0.88 0.89 0.86 0.84 0.82

ρ 0.01
LCR 0.89 0.89 0.88 0.86 0.87 0.82
Nants 10
(4) GM 0.95 0.90 0.91 0.90 0.88 0.87
Nmax 2500
WM 0.94 0.89 0.91 0.87 0.85 0.83
Q0 0.6
HACSEG γ 1 Table. 2. Performance comparison of the (1) SA, (2) GA,
(3)ACS-MRF, (4) HACSEG algorithms for segmentation of
β 2 MR images.
ρ 0.01
It is clear that the new algorithm HACSEG
Nants 10 outperforms others algorithms and finds better results.
Nmax 2500
Real images.
Table1. Parameters of the Simulated annealing, Genetic
algorithm, ACS-MRF and HACSEG algorithms. HACSEG algorithm has been applied to two gray
level real world images representing a muscle cell
where T0 : initial temperature, Tm : temperature image and a house image, presented in Figure 4.
multiplier, Ni : number of iterations after which
temperature is reduced, Nmax: maximum number of
iterations allowed, N the population size, Pc :
crossover probability, Pm : mutation probability, q0 :
parameter with determine the relative importance of
exploitation versus exploration, ρ, α : the pheromone
decay parameters, β : the relative importance of
pheromone trails, γ : the relative importance of the
heuristic function, and Nants : number of ants.
To validate the accuracy and reliability of the
segmentation method, compared with the ground a b
truth, we computed the Jaccard similarity. The Jaccard
similarity measures the similarity of two sets as the
Figure. 4. real world images. a) muscle cells image, b) a
ration of the size of their intersection divided by the house image
size of their union. Let Vgk and Vsk denotes the total
number of pixels labeled into a cluster k in the ground To quantitatively evaluate the quality of the
truth (g) and the obtained segmentation (s). For cluster segmentations, without need to a ground truth, we use
k
k the Jaccard similarity J (g,s) is defined by a simplified version of the Borsotti-measure (Borsotti
& al, 1998), which was defined as follows:
Vgk ∩Vsk
J (g , s )=
k
(9)
Vgk ∪Vsk
R  2 
e)
k Q (I ) = 1 R ∑ k 
A good segmentation is obtained when J (g,s) is 1000(N) k =1  A
 k 
near 1 which means that the cluster k is well detected.
(10)
All algorithms are run during 10 test runs but only where:
the best result is reported in table. 2. The results of I : image to be segmented;
other algorithms are obtained from (Ouadfel & al,
2003a). N : size of the image;
R : number of clusters in the segmented image;
N= 0% N = 3% N = 5%
SETIT2005

Ak : area, or the number of pixels of the kéme (Bonabeau & al, 1999) Bonabeau, E., Dorigo M. and
cluster; Therualez, G. “ Swarm Intelligence: From Natural to
ek : gray level error of cluster k. Artificial Systems”. New Work: Oxford University
Press, 1999.
(Borsotti & al, 1998) Borsotti, M., Campadelli, P. and
ek is defined as sum of the square of the Euclidean
Shettini, R. “Quantitative evaluation of color
distances between the gray level intensity of pixels in image segmentation results”, Pattern Recognition
cluster k and the average gray level intensity of cluster letters 741-747, 1998.
k in the segmented image.
(Bullnheimer & al, 1997) Bullnheimer, B., Hartl, R. and
The first term of Eq. (10) is a normalization factor. Strauss, C. “Applying the Ant System to the Vehicle
The smallest the value of Q(I), the better the Routing Problems”, 2nd Metaheuristic International
segmentation results were. Table 2 summeatizes the Conference (MIC-97), Sophia-Antipolis, France, 1997.
results obtained from SA, GA, ACS-MRF and (Chellappa & al, 1993) Chellappa, R. and Jain, A.K.
ACSEG algorithms. Editors, Markov Random Fields: Theory and
Images (a) (b) Application Academic Press, San Diego. 1993.

SA 1.130 0.980 (Chen & al, 1995) Chen, J.L. and Kundu, A.,
“Unsupervised texture segmentation using multichannel
GA 1.128 0.950
decomposition and hidden Markov-models”. IEEE
ACS-MRF 1.128 0.9486 Trans. Image Process. (4) pp. 603–619, 1995.
HACSEG 1.1272 0.9473 (Costa & al, 1997) Costa., D. and Hertz., A. “Ants Can
Color Graphs”, Journal of the Operational Research
Table. 2: Borsotti & al. Measure (Q) comparison for real
Society, (48): 295-305, 1997.
images (a) and (b)
(Di Caro & al, 1998) Di Caro, G. and Dorigo, M. "AntNet:
As we can see, HACSEG obtained good quality Distributed Stigmergetic Control for Communications
segmentation results. Networks", Journal of Artificial Intelligence Research
(JAIR), (9): 317-365, 1998.

6 Conclusion (Dorigo & al, 1991) Dorigo, M., Maniezzo, V. and Colorni,
A., “Positive feed back as a search strategy”. Technical
In this paper we have described HACSEG, a new Report (16). Politecnico di Milano, Italy. 1991.
ant algorithm for image segmentation based on a
modified version of the Ant Colony System (ACS) (Dorigo & al, 1997a) Dorigo, M. and Gambardella, L.,
algorithm coupled with a local search. Instead of “Ant Colony System: A Cooperative Learning Approach
to the Traveling Salesman Problem”. IEEE Transactions
building a new partition each time, the ant algorithm
on Evolutionary Computation (1), pp: 53-66, 1997.
starts with a number of random partitions. Each
partition is first optimized and then associated with (Dorigo & al, 1997b) Dorigo, M. and Gambardella, M.
one ant, which applies a given number of local moves “Ant Colony for the Traveling Salesman Problem”,
to pixels. The moves are chosen according to the BioSystems, (43): 73-81, 1997.
pheromone trails and heuristic information. A simple (Dorigo & al, 1999) Dorigo, M., Di Caro,G. and
local search algorithm is used to improve the quality Gambardella, L. “Ant Algorithms for Discrete
of the partition found by each ant and yielding a faster Optimization”, Artificial Life. 2(5): 137-172, 1999.
convergence of the algorithm. Experimental results
(Dubes & al, 1990) Dubes, R, Nadabar, S., Chen, C.C. and
show a noticeable increase in performance compared Jain, A.K., “MRF Model-Based algorithm for Image
to ACS-MRF algorithm and other global optimization Segmentation” Proc. 10th Int’l. Conference on Pattern
methods like SA and GA. Recognition, Atlantic City, pp. 808-814, 1990.
(Franti & al, 2000) Franti, A. and Kivijarvi, J.,
Références “Randomized Local Search Algorithm for the Clustering
(Andrey & al, 1998) Andrey, P. and Tarroux, P., Problem”. Pattern Analysis & Applications (3). 358-369,
“Unsupervised segmentation of Markov random field 2000.
modeled textured images using selectionist relaxation”. (Gambardella & al, 1997) Gambardella, L. and Dorigo, M.
IEEE Trans. Pattern Anal. Machine Intell. (20) 3, pp. ”HAS-SOP: An hybrid ant system for the sequential
252–262, 1998. ordering problem”, Technical Report (11), IDSIA,
(Bertelle & al, 2003) Bertelle, C., Dutot, A., Guinand, F., Lugano, CH. 1997.
Olivier, D. “Color Ant Populations Algorithm for (Gambardella & al, 1999) Gambardella, L., Taillard, E.,
Dynamic Distribution in Simulations”. In ESS 2003 and Dorigo, M. “Ant Colonies for the QAP”.
European Simulation Symposium, Delft (Holland) J.Oper.Res.Soc., 50,167-176. 1999.
October 2003.
(Geman & al, 1984) Geman, S. and Geman, D., “Stochastic
(Besag, 1986) Besag, J. “On the statistical analysis of dirty relaxation, Gibbs distribution, and Bayesian restoration
pictures”, J. Roy. Statist. Soc. B 3(48): 259–302, 1986. of images”. IEEE Trans. Pattern Anal. Machine
Intell.(6). pp. 721–741, 1984.
SETIT2005

(Gunnels & al, 1994) Gunnels, J., Cull, P. and Holloway, (Panjwani & al, 1995) Panjwani, D.K. and Healey, G.,
J.L. “Genetic algorithms and simulated annealing for “Markov random field models for unsupervised
gene mapping”, In: Proc. 1st IEEE Conf. Evolutionary segmentation of textured color images”. IEEE Trans.
Computation, 385–390. 1994. Pattern Anal. Machine Intell. (17) 10, pp. 939–954.,
1995.
(Hu & al, 1992) Lu, Y. and Dennis, T.J. “Simulated
annealing and iterated conditional modes with selective (Ramos & al, 2002) Ramos, V. and Merelo, J.J. Self-
and confidence enhanced update schemes”, In: Proc. 5th Organized stigmergic Document Maps:
Annual IEEE Symp. Computer-based Medical Systems, Environment as a Mechanism for Context
257–264, 1992. Learning, in AEB´2002 – 1st Spanish Conference on
Evolutionary and Bio-Inspired Algorithms, E. Alba, F.
(Huang & al, 1995) Huang, Y., Palaniappan, K., Zhuang X.
Herrera, J.J. Merelo et al. (Eds.), pp. 284-293, Centro
and Cavanaugh, J.E. “Optic flow field segmentation and
Univ. de Mérida, Mérida, Spain, 6-8 Feb. 2002.
motion estimation using a robust genetic partitioning
algorithm”, IEEE Tranactions and Pattern Analysis. (Solnon & al, 2004) Solnon, C., and Fenet, S.
Machine Intelligence. (17), pp. 1177–1190. 1995. ”Investigating ACO capabilities for solving the
maximum clique problem”. Research report LIRIS RR-
(Kato, 1994) Kato, Z. “Modélisation Markoviennes
2004-021, Journal of Heuristics / Kluwer - May 2004
Multiresolutions en vision par ordinateur. Application à
la segmentation d’images SPOT”, Thèse de l’Université (Taillard & al, 1997) Taillard, E. and Gambardella, L. “An
de Nice Sophia Antipolis, 1994. Ant Approach for Structured Quadratic Assignment
Problems”, 2nd Metaheuristic International Conference
(Kervrann & al, 1995) Kervrann, C. and Heitz, F., “A
(MIC-97), Sophia-Antipolis, France, 1997.
Markov random field model based approach to
unsupervised texture segmentation using local and (Trejos & al, 2004) Trejos, J., Murillo, A. and Piza, E.
global spatial statistics”. IEEE trans. Image Process. (4 “Clustering by Ant Colony Optimization''. In Meeting of
), 1995. the International Federation of Classification Societies,
2004.
(Kim & al, 1998) Kim, E. Y. , Park, S. H. and Kim, H. J. “A
Genetic algorithm-based segmentation of Markov .
Random Field modeled images”. IEEE Signal
processing letters 7 (11), pp. 301–303., 2000.
(Li, 1995) Li, S.Z. Markov Random Field Modeling in
Computer Vision, Springer, Tokyo, 1995.
(Lin & al, 1993) Lin, F.T., Kao, C.Y. and Hsu, C.C.
“Applying the genetic approach to simulated annealing
in solving some NP-hard problems”, IEEE Trans. Syst.
Man Cybernet. 6(23) : 1752–1767,1993.
(Maniezzo & al, 1994) Maniezzo, V., Colorni, A. and
Dorigo,M. “The Ant System Applied to the Quadratic
Assignment Problem”, Technical Report. IRIDIA. (28),
Université Libre de Bruxelles, Belgium, 1994.
(Maniezzo & al, 2001) Maniezzo, V., Carbonaro, A., “Ant
Colony Optimization: an overview”. In C.Ribeiro (eds.)
Essays and Surveys in Metaheuristics, Kluwer, page.21-
44, 2001.
(Meshoul & al, 2003) Meshoul, S. and Batouche, M.
“Robust Point Correspondence for Image
Registration using a colony of cooperating artificial
ants”. Pp 1102-1110, WSEAS Transactions on
systems, Issue 4, V(2), 2003.
(Ouadfel & al, 2003a) Ouadfel, S. and Batouche, M.
”MRF-based image segmentation using Ant
Colony System’. ELCVIA: Electronic Letters on
Computer Vision and Image Analysis 2(1):pp 12-24,
2003.
(Ouadfel & al, 2003b) Ouadfel, S., Batouche, M. and
Talhi, S. “MRF-based image segmentation using
autonomous agents”. International Conference:
Sciences of Electronic, Technologies of Information and
Telecommunications, SETIT 2004 TUNISIA
(Pal & al, 1993) Pal, N.R. and Pal, S.K., “A review on
image segmentation techniques”. Pattern Recognition
(26) 9, pp. 1277–1294., 1993.