IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO.

4, JUNE 2010 267

A Stochastic Approach to Image Retrieval Using

Relevance Feedback and Particle Swarm Optimization
Mattia Broilo, Student Member, IEEE, and Francesco G. B. De Natale, Senior Member, IEEE

Abstract—Understanding the subjective meaning of a visual of the user? Third, is there a reliable method to cluster relevant
query, by converting it into numerical parameters that can be ex- and irrelevant images, taking into account that, even if relevant
tracted and compared by a computer, is the paramount challenge images may luckily represent a compact cluster, irrelevant ones
in the field of intelligent image retrieval, also referred to as the
“semantic gap” problem. In this paper, an innovative approach for sure will not? Simple minimum-distance-based algorithms
is proposed that combines a relevance feedback (RF) approach are usually unable to provide a satisfactory answer to all such
with an evolutionary stochastic algorithm, called particle swarm problems.
optimizer (PSO), as a way to grasp user’s semantics through According to this, several additional mechanisms have been
optimized iterative learning. The retrieval uses human interac-
tion to achieve a twofold goal: 1) to guide the swarm particles introduced to achieve better performance. Among them, rele-
in the exploration of the solution space towards the cluster of vance feedback (RF) proved to be a powerful tool to iteratively
relevant images; 2) to dynamically modify the feature space by collect information from the user and transform it into a se-
appropriately weighting the descriptive features according to the mantic bias in the retrieval process [4]. RF increases the re-
users’ perception of relevance. Extensive simulations showed that
trieval performance thanks to the fact that it enables the system
the proposed technique outperforms traditional deterministic
RF approaches of the same class, thanks to its stochastic nature, to learn what is relevant or irrelevant to the user across succes-
which allows a better exploration of complex, nonlinear, and sive retrieval-feedback cycles. Nevertheless, RF approaches so
highly-dimensional solution spaces. far proposed show some critical issues yet unsolved. First, user
Index Terms—Content-based image retrieval, particle swarm interaction is time consuming and tiring, and it is therefore de-
optimizer (PSO), relevance feedback (RF). sirable to reduce as much as possible the number of iterations to
convergence. This is particularly difficult when only a few new
images (possibly none) are retrieved during the first RF steps, so
I. INTRODUCTION that no positive examples are available for successive retrieval
ONTENT-BASED image retrieval (CBIR) systems ana-
C lyze the visual content description to organize and find
images in databases. The retrieval process usually relies on pre-
[5]. In this case, the method should introduce some alternative
strategy to explore the solution space (e.g., some perturbation
of the solution).
senting a visual query (natural or synthetic) to the systems, and Another critical problem concerns the risk of stagnation,
extracting from a database the set of images that best fit the user where the search process converges to a very suboptimal local
request. Such mechanism, referred to as query-by-example, re- solution, thus being unable to further explore the image space.
quires the definition of an image representation (a set of descrip- This problem is more and more evident when the size of the
tive features) and of some similarity metrics to compare query database increases. Again, additional mechanisms to allow
and target images. enlarging the exploration are usually needed.
Several years of research in this field [1]–[3] highlighted a In order to overcome the above problems, in this paper, we in-
number of problems related to this (apparently simple) process. vestigate the possibility of embedding the RF process into a sto-
First, how good is the description provided by the adopted fea- chastic optimization engine able to provide on one side a better
ture set, i.e., are the selected features able to provide a good clus- exploration of the search space, and on the other side to avoid
tering of the requested images, retrieving a sufficient number the stagnation in local minima. We selected a particle swarm
of desired images and avoiding false positives? Second, is the optimizer [6], for it provides not only a powerful optimization
query significant enough to represent the conceptual image that tool, but also an effective space exploration mechanism. We
the user has in mind, i.e., does the query capture the semantics would like to point out that the optimal choice of the features
used for image description is outside the scope of this paper,
Manuscript received May 19, 2009; revised September 22, 2009; accepted and then all the tests presented are based on a very standard
November 16, 2009. First published March 22, 2010; current version published description based on colors and textures. A very preliminary
May 14, 2010. A preliminary version of this work has been previously presented
at 10th Workshop on Multimedia Signal Processing-MMSP, Cairns, Australia,
version of this work was presented in [7], where the concept
2008. This work was supported in part by the EU Commission under the frame- of PSO-CBIR is first introduced. Besides introducing a deeper
work of the LivingKnowledge project (EU FET Grant no. 231126). The asso- description and experimental analysis of the basic method, the
ciate editor coordinating the review of this manuscript and approving it for pub-
lication was Dr. Zhu Liu.
present work presents a new schema, where we introduce the
The authors are with the Department of Information Engineering and concept of swarm evolution with iterative swarm subdivision to
Computer Science (DISI), University of Trento, 38050 Trento, Italy (e-mail: allow a much more extensive exploration of the solution space
[email protected]; [email protected]). and a faster and better convergence.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. The paper is organized as follows: in Section II, a thorough
Digital Object Identifier 10.1109/TMM.2010.2046269 explanation of the motivations of this work is proposed, pointing
1520-9210/$26.00 © 2010 IEEE

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
268 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO. 4, JUNE 2010

out the choice of relevance feedback and the effectiveness of technique that allows solving complex optimization problems
stochastic approaches in the image retrieval task. Section III [22]. It is inspired by the behavior of swarms of bees, where a
gets deeper on the state of the art concerning RF-based image particle corresponds to a single bee that flies inside a problem
retrieval and PSO algorithm. Section IV introduces a detailed solution space searching for the optimal solution. During the
description of the proposed PSO-based relevance feedback ap- iterations, the particles move towards the swarm’s global best
proach for CBIR. In particular, the RF mechanism and the cus- and the particle’s personal best, which are known positions
tomization of the PSO to the specific problem are addressed, in the solution space. These positions represent the social and
stressing the peculiarities of the proposed implementation with the cognitive factor of the swarm and are weighted by two
respect to generic PSO implementations. Section V describes parameters that influence the swarm behavior in the search
the experimental set up, specifying the adopted feature sets and process.
the test datasets, while Section VI reports a set of representative PSO has been successfully applied as an optimization tool in
results, paying particular attention to parameter tuning and per- several practical problems [23] and in many different domains
formance evaluation. Finally, Section VII draws the conclusions such as image classification [24], ad-hoc sensor network [25],
and identifies future research directions. design of antennas array [26], and neural networks [27]. PSO
has been used in many cases as a way to generate optimized
II. BACKGROUND parameters for other algorithms. As such, it has been proposed
also in the field of CBIR. In particular, in [28] and [29], it is used
A. Relevance Feedback to build a supervised classifier based on self-organizing features
maps (SOM), while in [30], it is applied to the tuning of param-
As already mentioned, the proposed technique is based on the eters of a similarity evaluation algorithm. An in-depth study on
well-known concept of relevance feedback. The basic RF mech- the use of statistical methods in image retrieval problem was
anism consists in iteratively asking the user to discriminate be- recently done by Chandramouli and Izquierdo [31], where the
tween relevant and irrelevant images on a given set of results [8]. image retrieval task is treated as a classification problem.
The collected feedback is then used to drive different adaptation
mechanisms which aim at better separating the relevant image
cluster or at reformulating the query based on the additional user III. MOTIVATIONS
input. In the first case, we may apply feature re-weighting [9] Although RF is for sure not new in itself, no viable alter-
or adaptation [10] algorithms, which modify the solution space native solutions have been proposed so far to capture the user
metrics, giving more importance to some features with respect semantics in content-based image retrieval through user inter-
to others. In the second case, also known as query shifting, we action, unless additional information is available (e.g., textual
will move the initial user query towards the center of the rele- keywords, such as tags or annotations). RF has been used in dif-
vant image cluster to obtain a new “virtual query”, which takes ferent fields of information retrieval, but its current moderate
into account the multiple inputs of the user across iterations [11]. success in the media domain is mainly due to the limited per-
Feature re-weighting and query shifting are often used jointly. A formance achieved by available algorithms, which require nu-
binary RF is used to train neural network systems as in PicSOM merous iterations to achieve a significant number of relevant im-
[12] or in the work of Bordogna and Pasi [13]. In [14], a fuzzy ages. As a matter of fact, RF mechanisms currently used in some
RF is described, where the user provides the system with a fuzzy beta or demo version of online search engines usually rely on a
judgment about the relevance of the images. It was also pro- simple substitution of the query with one of the images found in
posed to exploit a RF approach to model an SVM-based clas- the previous search. In this case, the history of the search is not
sifier: this is the case of the work by Djordjevic and Izquierdo maintained, making impossible to achieve a real adaptation of
[15], and of the system developed by Tian et al. [16]. A thor- the search. It is our opinion that more sophisticated methods are
ough survey on the existing RF techniques for image retrieval is likely to be adopted in the future if effective methods to exploit
presented in [17], while two different evaluations and compar- the history of the search will be available.
isons of several RF methods and schemas are reported in [18] The proposed method converts the RF into an optimization
and [19]. algorithm, thus opening a new perspective for the development
of more efficient and computationally-effective RF approaches.
B. PSO Optimization
To this purpose, we restate the problem of finding the images
In the last years, the development of optimization algorithms that match a given user query as an optimization problem where
has been inspired and influenced by natural and biological the requested images are the ones that minimize a given fitness
behaviors [20]. Bio-inspired optimization approaches provided function. A swarm particles fly in a multidimensional feature
new ways to achieve nearly-optimal solutions in highly non- space populated by the database images. The features provide
linear, multidimensional solution spaces, with lower complexity the image description, and each image is uniquely represented
and faster convergence than traditional algorithms. In this paper, by a feature vector, corresponding to a point in the space. The fit-
we investigate the use of a popular bio-inspired stochastic op- ness function is defined such that minimum values are achieved
timization algorithm called particle swarm optimization (PSO) when particles approach images which fit the user’s request.
to achieve an efficient interactive CBIR algorithm. PSO was in- Then, swarm migration process is run so that particles may itera-
troduced in the field of computational intelligence by Kennedy tively converge to the solution that minimizes the global fitness,
and Eberhart [21] in 1995 and is a population-based stochastic i.e., to the cluster of images that best fit the user query.

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
BROILO AND DE NATALE: STOCHASTIC APPROACH TO IMAGE RETRIEVAL 269

and the distance between query and image is calculated as a

weighted Euclidean distance computed among feature vector
pairs [33]. Initially, the weights are all equal to 1. Then, the
nearest images are presented to the user, and the first feedback
is requested. The feedback is binary, and labels each retrieved
image as relevant or irrelevant. Accordingly, two image subsets
are created, which will be progressively populated across itera-
tions thanks to human interaction.
The definition of relevant and irrelevant image subsets makes
it possible to perform a first re-weighting of the features and a
first updating of the swarm. Details about such procedures are
provided in Section IV-A. After that, a new ranking is calcu-
lated based on the weighted Euclidean distance (with updated
weights) and the nearest images are again proposed to the
user to collect a new feedback. The process is then iterated until
convergence.
During this process, the feature weights are iteratively speci-
fied to fit the user’s mental idea of the query, i.e., the two clas-
sified image subsets allow the system to understand which fea-
tures are more important to discriminate between relevant and
irrelevant images. In parallel, the optimization process is car-
Fig. 1. Flowchart of the proposed approach. ried out by constantly updating the swarm, which progressively
converges to the image cluster that contains the best solutions
found across iterations.
We will demonstrate that using swarm intelligence of the PSO
algorithm, it is possible to substitute a generic query shifting A. Query Selection and Dist Calculation
[32] by using completely different process, where the particles
of the swarm can be seen as many single retrieval queries that The first operation is to describe the images in terms of fea-
search in parallel, locally and globally, moving towards relevant tures. The visual signature of the th image is made of four
samples and far from irrelevant ones. Practically speaking, this different feature vectors, composed by: color moments
can be seen as a generalized query shifting algorithm, where , color histogram bins , edge histogram bins
each particle of the swarm can be considered as a query point , and wavelet texture energy values . The vector
that searches in parallel the best solution inside a local area of of dimension
the feature space, and the different queries (particles) are com- provides then the overall description of the image.
bined by taking into account the global and the personal best. The computation of the parameters is usually performed offline
A further added value of the proposed PSO-RF with respect to for database images and online for the user query. From that
other proposed CBIR algorithms is in the fact that it introduces point on, each image is completely represented by its visual sig-
a stochastic component to the process, thus allowing to explore nature, or equivalently by a point in an -dimensional space.
the solution space in different ways, thus making it possible to After the selection of the query and its mapping in the feature
climb local minima and to converge to a good solution inde- space, the system shows the user the most similar images
pendently of the starting point and of the path followed. This is in the entire database according to (1):
achieved by making three components cooperate in the conver-
gence process with appropriate weights: a social factor (where
the swarm did find the best solution), a cognitive factor (where
the particle did find its best solution), and an inertial factor (to-
wards where the particle is moving).
(1)
IV. PROPOSED PSO-RF APPROACH where WMSE is the weighted Euclidean distance calculated be-
In this section, we describe the proposed retrieval algorithm tween a pair of corresponding feature vectors:
that we will call PSO-RF. PSO-RF is based on two iterative
processes: feature re-weighing and the swarm updating. Both (2)
processes use the information gathered from the user, who is
iteratively involved in the image search process. Fig. 1 shows
the block diagram of the proposed algorithm. According to the where and are two generic feature vectors, is a vector
classic “query-by-example” approach, the user selects the query of weights associated to the features ( , where is
image, and based on that, the system ranks the whole dataset equal to or or or ), and marks the iteration
according to a minimum distance criterion. To this purpose, number. At the first iteration , all the features are equally
each image (included the query) is mapped into a feature vector important, i.e., ; .

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
270 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO. 4, JUNE 2010

After computing ; ; , where One of the most important points in an optimization process
is the number of database images, a ranking is performed is the definition of the target function to be minimized or max-
to sort the database according to the distance from the query. imized, called fitness. The fitness function should represent the
Then, the ranked list is shown to the user to collect the feedback. effectiveness of the solution reached by the swarm particles.
Taking into account the relevant and irrelevant image sets, (5)
B. User Feedback and Features Reweighting defines the weighed cost function that expresses the fit-
ness associated to the solution found by the generic particle :
The above metric provides a quantitative measure of the vi-
sual dissimilarity of two images. It is then used to compare each
of the images in the database with the query, and to sort
them in increasing distance order. After that, the first re-
sults are shown to the user to collect the relevance feedback. In
particular, the user is asked to tag the presented images as (5)
relevant or irrelevant according to his mental idea of query. Two
image subsets are then created, namely relevant and irrel- where ; and ; are
evant sets. From this point on, the two sets are maintained the images in the relevant and irrelevant image subsets, respec-
and updated during all the iterations, preserving the history of tively. The weight vector needed to compute is the one
the retrieval process. calculated at the previous step. It is to be observed that the func-
The aim of weight updating is to emphasize the most discrim- tion produces lower values when the particle is close
inating parameters. In practice, the idea is to perform a dynamic to the relevant set and far from the irrelevant one. Therefore,
feature selection, driven by the user feedbacks (used as a su- the lower the fitness, the better the position of the particle. Ac-
pervision input). The feature re-weighing algorithm used in this cording to the fitness value, it is possible to reorder the particle
work is similar to the one proposed in [34], and is based on a swarm obtaining a new ranking. It is also worth noting that both
set of statistical characteristics. Taking into account the concept the weights and the fitness function change across iterations be-
of dominant range (that is the range of a single feature of the cause of the dynamic nature of and subsets. Ac-
image subset ), it is possible to calculate the discriminant cordingly, features that were relevant to discriminate some im-
ratio on the th feature , at the th iteration, ages can lose importance and particles that were considered very
which indicates the ability of this feature to separate relevant im- close to the global best can become far from the relevant zone
ages from irrelevant ones: of the solution space. In most cases, the number of irrelevant
images collected across iterations is greater than the number
of relevant ones; this aspect has been taken into consideration
(3) during the formulation of the fitness function, making it depen-
dent on the inverse of the distance from irrelevant images. In
where is the number of irrelevant images at the th iter- this way, the more the average distance of the particle from ir-
ation, while is the number of images in that have relevant images grows, the more the fitness depends only from
the feature within the range associated to the corresponding relevant images.
feature in . The weights are then updated as follows:
D. Evolution and Termination Criteria
Having defined all the elements of the optimization process,
(4)
we still need to identify how to make the swarm evolve in time.
To do that, we have to define some attributes of the particles.
where is the standard deviation of the th feature in Each particle holds a personal best (a relevant position)
at the th iteration. To avoid problems when is close to and a global best that is the best position among all the solu-
zero, the method implemented in [34] has been modified with a tions found during the whole retrieval process (the same position
normalization factor that limits the maximum weight to 1. for all the swarm particles).
In our approach, the selection and the update of personal
C. Swarm Initialization and Fitness Evaluation (pbest) and global best (gbest) are very different from a typical
PSO implementation [35]. The global best is updated at each
As previously mentioned, in our work the retrieval is formu- iteration as an image of the “relevant” set . The image is
lated as an optimization process. We have therefore to model selected according to (6): for each relevant image, the sum of
the retrieval problem in terms of a PSO. To this purpose, we de- the distances from the other relevant images is calculated; then,
fine the swarm particles as points inside the feature space, the image resulting nearest to the others is chosen as gbest. If
i.e., an -dimensional vectors in the feature space. Given a there are no relevant images except for the query, gbest remains
number of particles with , we initialize the the position of the query:
swarm by associating each particle to one the nearest neigh-
bors of the original query, according to the ranking performed
at first iteration. Then, we generate a random speed vector ;
independently for each swarm particle, to ini-
tialize the stochastic exploration. (6)

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
BROILO AND DE NATALE: STOCHASTIC APPROACH TO IMAGE RETRIEVAL 271

called acceleration coefficients, aimed at pulling the particle to-

wards the position related to the cognitive (i.e., personal best )
or social part (i.e., global best ). Further details on the PSO
parameter choice are reported in Section V. Finally, the position
of each particle is updated as follows:

(8)

where the sign of the speed is changed according to the “re-

flecting wall” boundary condition in order to limit the search
of the relevant images inside the space of admissible solutions
[37]. It is worth noting that it is possible to view the particles of
the swarm like query points that will explore the -dimensional
solutions space, made of the image features , with
an own speed and direction. It is useful to point out that the im-
ages of the database represent a discrete
and fixed set of points, while the particles can move in a contin-
uous way inside the features space.
After the swarm initialization at first iteration (setting the ini-
tial position, the random speed, the pbest, and the gbest), an up-
dating operation is done at every following iteration. The gbest
image is recalculated if new relevant images are tagged by the
user, and the new speed and position of each particle are calcu-
lated according to (7) and (8), respectively. Consequently, after
every user feedback, the swarm moves towards new areas in the
solution space where other relevant images may be found.
To complete a single iteration, a further operation is needed:
to associate to particles placed in a “good position” (the lower
the fitness, the better the position of the particle), the nearest
Fig. 2. Example of gbest and pbest evolution.
images in the database according to (1). In fact, the particles
move semi-randomly in a continuous space, while database im-
ages are in discrete positions. Then, the first particles of
pbest is different for each particle and is initialized as the feature the swarm ranked according to (5) are associated to their closest
vector originally associated to the particle. Until the retrieval image, thus obtaining the new set of images to be presented to
algorithm does not find any relevant images, pbest will be up- the user. If more than one particle points to the same image or
dated according to the fitness function (5) only if a particle points to an image already classified as irrelevant, the
. If the user tags some retrieved images as relevant, corresponding image is discarded and next nearest neighbor is
the swarm will be split into sub-swarms. While gbest position considered, until different images are collected. After the
(relevant centroid) is shared among all the sub-swarms to guar- user feedback, the above-described process of re-weighting and
antee the continuity of the convergence process, pbest is forced swarm updating is iterated. The process ends when one of the
in the position of the relevant images. In this way, it is possible to following conditions is verified: 1) the user is satisfied with the
better explore the features space near the relevant points, while result of the search, 2) a target number of relevant images is
maintaining a global reference point. An example of these con- achieved (in general ), or 3) a predefined number of iter-
cepts is provided in Fig. 2. ations is reached. After termination, all relevant images found
In our tests, the initial swarm evolves splitting itself until a are shown to the user.
maximum of sub-swarms. Using the knowledge of the
global and individual best, the speed of each particle is set, ac- Remarks on the Optimization Strategy
cording to the following (7):
It has to be pointed out that our use of PSO substantially dif-
fers from previous works in the same field. First, we have no
(7) complete knowledge about the problem, e.g., the class “irrel-
evant images” is highly variable, and the supervision informa-
where and are uniform random numbers in the range ; tion is very limited and not completely available from the begin-
and is an inertial weight parameter in , progressively ning. Second, the representation (description) of our objects is
decreasing along iterations [36]. largely uncorrelated with the classification of the objects them-
The inertial weight is calculated in such a way to decrease selves (from where, the semantic gap). Finally, we do not aim
proportionally to the number of retrieved relevant images, thus at finding an optimal point in the solution space, but we want to
allowing to slow down the swarm when approaching to conver- use the swarm as a way to explore a feature space to find the best
gence. The parameters and are two positive constants matching points, according to a cost function to be optimized.

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
272 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO. 4, JUNE 2010

Then, the maximization of the fitness becomes a side effect (al- Of course, these assumptions may lead to results which do not
though fundamental to achieve the solution) of the convergence correspond to the subjective behavior of a generic user. In fact,
of the swarm to the set of relevant images. an image may encompass several coexisting visual concepts or
Furthermore, the basic PSO scheme is modified by intro- even in the same object. As an example, the image class “cats”
ducing a “divide and conquer” schema, where the swarm can be can be included at large in the image class “animals”, or be fur-
split according to the number of relevant images retrieved. We ther specified in a subclass “black cats”. At the same time, an
will show that this procedure has a twofold goal: to avoid stag- image containing a cat may be classified according to another
nation, and to allow the convergence of the swarm to multiple subject present in the image, e.g., a dog, which is considered the
sub-clusters where relevant images are. As to the first point, we main subject by the user. Additionally, an image can be classi-
have to consider that we could have only a very limited number fied according to some abstract concept connected to the subject
of iterations, and just very partial information on the ground (e.g., the activity it is performing, or the way it is behaving),
truth (the initial query and the images labeled by the user across making it even more difficult to ensure a significant labeling. In
iterations). If the number of relevant images retrieved in the be- all those cases, a relevant image can be classified as irrelevant
ginning of the process is low, the risk of stagnation is very high (or vice-versa) during the process and in the final evaluation,
[7]. The adopted methodology sharply reduces such problem, thus preventing the convergence and in any case showing sub-
increasing the exploration capabilities of the algorithm. As to optimal performance. Several studies are being carried on how
the second point, we observed that, being the representation of to manage these problems through the use of appropriate knowl-
images just based on generic visual features while classes are edge (for instance, tagging, taxonomies, and ontology), which
usually organized on the basis of visual concepts, often relevant are beyond the scope of this paper.
images are not grouped in a single cluster in the feature space. In order to limit such problems, we were very careful in
Therefore, sub-swarms are more suited to handle this problem. checking the consistency of the dataset we used for testing. Such
Finally, another important difference from typical optimiza- datasets were achieved by merging two different and widely
tion problems is that the data are not completely available since used databases, and selecting the largest possible number of
the beginning, but they are collected from the user feedback image categories that presented a sufficient consistency. The
across iterations. This “online supervision” makes the conver- selected databases were the Caltech-256 [38] and the Corel
gence faster (as compared to standard implementations), since Photo Galleries (http://www.corel.com), for their large use in
the learning procedure is directly driven by the user knowledge. the scientific literature (see, e.g., [39] and [40]). The resulting
dataset includes 150 categories, each one represented by 80 im-
V. EXPERIMENTAL SETUP ages, for a total of 12 000 images. Of course, this may turn out
to be limited as compared to the huge number of images used
In this section, we provide some details about the setup of in large-scale applications. However, our objective here is to
the tests described and commented in Section VI. In particular, demonstrate that our method compares favorably to previously
we illustrate the databases used for experimental testing, the proposed approaches, on the same type of dataset used thereby.
feature set adopted for image description, and the setting of PSO Additionally, a user interface has been developed to allow
parameters. user-oriented tests and a demo of the retrieval system. Further
experiments to attain subjective analyses about user satisfaction
A. Image Databases and Image Classification are being currently carried on. A simple demo has been made
At present, common standards and universally accepted data available at the following URL: http://mmlab.science.unitn.it/
corpora to assess the performance of image retrieval systems demos/.
are not available. Furthermore, it is commonly accepted that
introducing a subjective feedback in the testing of RF systems B. Visual Signature
makes it extremely difficult to evaluate the performances and The selection of a significant set of descriptive features is cru-
to make comparisons with other state-of-the-art approaches. In cial in CBIR. Specific retrieval problems and different appli-
fact, the user can change his mind during the retrieval, make cation domains may require a careful selection of the features
errors, and can give a personal interpretation to the data. This that best describe the image database, such as colors, textures,
last problem is even more relevant in our approach, due to the contours, shapes, etc. [41], [42]. As previously mentioned, an
stochastic nature of the algorithm, which generates at each run in-depth analysis about image descriptors is out of the scope of
different results across iterations, thus making possible for a real this paper. Thus, a quite common set of descriptors was adopted
user to follow different paths to the solution. based on low-level visual features selected according to current
Consequently, we made two major hypotheses: 1) we adopted multimedia standards. Such features well adapt to photographic
two commonly used databases that, although limited in the picture retrieval, and our goal will be to demonstrate that the pro-
image variety and number, provide a trustable pre-classifi- posed approach provides a better performance than other com-
cation of images and allow an easier comparison with other peting methods given an equal description.
state-of-the-art methods; 2) we adopted the usual procedure of As usual, feature extraction was performed offline, but for the
providing automatic feedbacks based on pre-classified datasets, query image, thus affecting to a negligible extent the retrieval
widely used in the literature. Also in this case, this choice guar- speed. The feature vectors associated to each image were stored
antees significant results in comparative evaluations, thanks to in a database for runtime access. The size of the feature vector
a consistent use of data and classification criteria. was set to 75, and specifically: color histogram bins,

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
BROILO AND DE NATALE: STOCHASTIC APPROACH TO IMAGE RETRIEVAL 273

color moments, edge histogram bins, and

wavelet texture energy values [43]. Color moments
included first-, second-, and third-order central moments of each
color channel in the HVS color space (mean value, standard de-
viation, and skewness, respectively). The color histogram is cal-
culated in the RGB color space, while the edge histogram is ob-
tained dividing the image into four parts and the edges into four
main orientations. Finally, the 18 texture features represent the
coarse, middle, and fine level frequency content of the image in
the wavelet domain. Features are normalized in the range Fig. 3. Example of two different retrieval runs with the same query.
according to their variance [44].

VI. RESULTS iteration. Due to the stochastic nature of the PSO-based algo-
rithms, all precision and recall values plotted for those algo-
In this section, a selection of experimental results is presented
rithms are obtained by averaging five consecutive runs for each
to demonstrate the effectiveness of the proposed approach. As
query. An example of two different runs using the same query
far as parameter tuning is concerned, PSO has a few operating
is showed in Fig. 3. The retrieved images are different both in
parameters, and we will show that its performance is very stable
amount and in retrieved temporal order.
across a large set of experiments for a fixed setting.
For all our experiments, we set the number of feedback im- B. Parameters Tuning
ages (constant across iterations) which is a conve-
nient number to avoid confusing or bothering the user during the PSO optimization is ruled by four parameters: the inertial
interactive phase. The performance of the system was assessed weight , the cognitive and social acceleration constants and
in terms of precision (number of retrieved relevant images over , and the swarm dimension . Such parameters have a unique
) and recall (percentage of relevant images retrieved across goal, i.e., to determine the trade-off between global and local ex-
all iterations with respect to the number of class samples). The ploration. The inertial term spins the particles off to explore new
precision-recall curves are calculated after ten iterations, but ad- areas of the solution space, while the cognitive and social terms
ditional charts are provided to show the convergence across iter- attract the solution toward previously found good points (per-
ations. All the precision-recall curves are calculated considering sonal and global, respectively). The dimension of the swarm is
the first 80 ranked images (corresponding to the number of im- typically related to the dimensionality of the problem (in partic-
ages per category in the dataset). ular, the dimension of the solution space), so that the number
of “agents” exploring the space is significant with respect to
A. Comparison Methods the extension of the space to be explored. Some restrictions on
and values have been defined by Kennedy [45], based
The experiments presented in this section are organized as on the swarm convergence. Such restrictions force particles to
follows. First we analyze the impact of different settings of the avoid escalating oscillatory behaviors, this can be achieved by
PSO-RF parameters on retrieval performance. Then, using the imposing
best configuration, we provide the global performance on the se-
lected databases. The charts will also provide a comparison be- (9)
tween our Evolutionary-PSO-RF method and three different re-
trieval algorithms. The first one is an earlier version of PSO-RF The chart in Fig. 4 provides an experimental confirmation of
we proposed in [7], using different fitness and distance func- this rule, where different combinations of and are tested
tions, and no swarm splitting. Furthermore, in that algorithm, with , providing similar results, while the two tested
gbest and pbest were calculated as positions in the feature space, combinations that exceed the threshold have a dramatic loss in
which usually do not correspond to real images. The other two performance. Best results are achieved when , i.e.,
methods are deterministic, and derive from a traditional query local and global attraction coincide. The curves are averaged
shifting method based on the Rocchio equation [8]. The first de- over 2500 queries.
terministic method uses a single query, and a setting of Rocchio As far as the inertial term is concerned, it determines which
equation parameter as follows: ; ; and , percentage of the particle velocity at the previous iteration is
thus stressing the importance of the initial query and of the rel- transmitted to the current one, and it is fundamental for con-
evant images found. The second exploits the same evolution vergence rate. Its value typically has to decrease progressively
principle of our method, creating a new query for each relevant to ensure a larger exploration in the beginning of the process
image found. To achieve a fair comparison, all of the four com- and a better convergence in the end. Several studies are reported
pared CBIR systems use the same feature re-weighting function on the setting of the inertial weight to ensure the swarm con-
[34], except for normalization, which is set in the range for vergence. In [46], Eberhart and Shi suggested to vary the iner-
all methods. Furthermore, all compared techniques exploit both tial weight linearly from 0.7 to 0.4 along iterations, to achieve
relevant and irrelevant images tagged by the user. Finally, they a large-scale exploration in the early stages of the algorithm,
use the same image similarity metric (1), which is also used by and a refined exploration of the local basin at convergence [47].
the two deterministic algorithms to rank the database at each In our case, we used an initial inertia of 0.7 and we decrease

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
274 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO. 4, JUNE 2010

Fig. 4. C1 and C2 calibration at the end of the tenth iteration: 2500 queries, Fig. 6. Particle number calibration at the end of the tenth iteration: 2500
linear decreasing inertia, swarm with 500 particles. queries, C 1 = C 2 = 2:0, decreasing inertial weight '.

Fig. 5. Inertial weight ' calibration at the end of the tenth iteration: 2500
queries, C 1 = C 2 = 2:0, swarm with 500 particles.
Fig. 7. Sub-swarms number calibration at the end of the tenth iteration: 2500
queries, swarm with 500 particles, C 1 = C 2 = 2:0, decreasing inertial
weight '.

it until 0.2 depending on the number of the relevant retrieved

images. If no relevant examples are found, the inertial weight size setting. Among them, the problem statement with its con-
remains constant to avoid stagnation; otherwise, the inertia de- sequent fitness formulation and the number of iterations are the
creases in order to slow down the swarm speed, thus allowing most relevant. In many cases, a large swarm allows covering
a better local search when approaching the convergence. Fig. 5 more easily large search spaces, with lower sensitivity to local
shows the results of some studies we performed on the setting minima. On the other side, it may create convergence problems,
of . The resulting performances are quite similar, except for and increases storage and computation needs [48]. Fig. 6 shows
the case when a large constant value is maintained until the end the results of some tests performed varying the swarm size in
of the process, thus preventing the convergence. The best solu- the range ; 16 being the number of feedback images
tion is achieved when the inertia decreases proportionally to the and 1000 being around 10% of the database size. It is pos-
number of retrieved relevant images. Also in this case, the re- sible to notice that the performance grows with the swarm size,
sults are averaged over 2500 examples. but there is a saturation value after which there is no further im-
provement (corresponding to around 100 particles).
C. Swarm Evolution
To underline the great potential of the swarm evolution in
A further consideration concerns the impact of the swarm the algorithm, a representative precision-recall test is reported
size and of the splitting process on the retrieval performances. in Fig. 7. During iterations, the swarm splits into sub-swarms
During years, empirical results have shown that the swarm size to better explore the solution space. The chart clearly shows
can influence to some extent the PSO performance, but no gen- the performance increase from single swarm to multiple sub-
eral rule has been found to optimize the number of particles for swarms, and also in this case, a saturation parameter is clearly
every problem. Many factors contribute to the optimal swarm identifiable when the number of sub-swarms approaches .

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
BROILO AND DE NATALE: STOCHASTIC APPROACH TO IMAGE RETRIEVAL 275

Fig. 8. Precision-Recall improvement graph during the iterations using 2500 Fig. 10. Precision-Recall comparison of the four methods considered at the end
query images. of the tenth iteration.

Fig. 9. Average precision comparison of the four methods considered during Fig. 11. Precision-Recall graph of image classes coming from different
ten iterations. databases at the end of the tenth iteration.

Another important factor for the viability of RF procedures are rapidly trapped by local minima, while the proposed one
concerns the capability of reaching a sufficient number of re- grows much faster and continues to grow for some more itera-
trieved images with a limited number of feedbacks. To give an tions. Fig. 10 shows the precision-recall curve of the same ex-
idea of the results achieved iteration by iteration, Fig. 8 plots the periment. From these two graphs it is possible to note that Evo-
precision/recall graphs across iterations, calculated each time on lutionary-PSO clearly outperforms the other three approaches.
the basis of the 80 best-ranking images at that stage. Analyzing It is to be pointed out that our dataset was built as a combina-
that chart, it is possible to see how the swarm evolves rapidly tion of two test databases, as explained in Section V. This makes
from the second iteration (the first is deterministic) to reach an the test particularly realistic and challenging, since the image
asymptote after six to seven iterations. For instance, after five types and the relevant categories are not homogeneous. Being
iterations, the user has been presented less than 80 images (due the PSO-RF a stochastic algorithm, one is not guaranteed that
to possible multiple presentation of the same relevant images), successive runs of the process (even with the same query) will
and is able to retrieve in the average 25 relevant images among produce equal results (in terms of retrieved images and order
the first 80 ranked in the whole dataset. of retrieval). For sure, starting from different query images, the
result is different. What was observed to be very stable is the
D. Final Evaluations and Comparisons convergence, i.e., the number of retrieved images at the end of
A comprehensive comparative test is finally presented that the process for the same image class. To show this fact, we in-
uses a total of 2500 queries selected from 150 classes (20 im- troduced a chart (Fig. 11) that plots the statistics for a reduced
ages per class). The chart in Fig. 9 compares the average per- set of representative classes. Each curve is calculated using in
centage precision of the four methods for increasing iteration turn 60 images of the same class as queries, to demonstrate the
number. It is possible to observe that the compared methods low sensitivity to a bad starting point. The tests referring to the

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
276 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO. 4, JUNE 2010

Corel archive typically perform better, in particular for some [2] A. W. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain, “Con-
image classes commonly used in the literature. On the other tent-based image retrieval at the end of the early years,” IEEE Trans.
Pattern Anal. Mach. Intell., vol. 22, no. 12, pp. 1349–1380, Dec. 2000.
side, the Caltech database contains some particularly difficult [3] R. Datta, D. Joshi, J. Li, and J. Z. Wang, “Image retrieval: Ideas, influ-
categories, where the performance drops dramatically, mainly ences, and trends of the new age,” ACM Comput. Surv., vol. 40, no. 2,
due to the fact that the automatic mechanism used for feedback Apr. 2008, article 5.
[4] Y. Rui, T. S. Huang, M. Ortega, and S. Mehrotra, “Relevance feedback:
and match analysis requires a very well-defined classification A power tool in interactive content-based image retrieval,” IEEE Trans.
of the database, which is not the case for some Caltech database Circuits Syst. Video Technol., vol. 8, no. 5, pp. 644–655, Sep. 1998.
classes. Furthermore, some particularly difficult classes, such [5] T. S. Huang, C. K. Dagli, S. Rajaram, E. Y. Chang, M. I. Mandel, G. E.
Poliner, and D. P. W. Ellis, “Active learning for interactive multimedia
as starfish, would require much more complex descriptive fea- retrieval,” Proc. IEEE, vol. 96, no. 4, pp. 648–667, Apr. 2008.
tures to be identified than those used in our tests. Nevertheless, [6] J. Kennedy, R. C. Eberhart, and Y. Shi, Swarm Intelligence. San
it is important to point out that also in these cases, the proposed Francisco, CA: Morgan Kaufmann, 2001.
[7] M. Broilo, P. Rocca, and F. G. B. De Natale, “Content-based image
method behaves better than the compared ones. retrieval by a semi-supervised particle swarm optimization,” in Proc.
One last important consideration concerns the convergence of Int. Workshop Multimedia Signal Processing, 2008.
the RF, and therefore the complexity of the proposed approach. [8] C. J. Van Rijsbergen, Information Retrieval, 2nd ed. London, U.K.:
Butterworths, 1979.
The chart in Fig. 7 clearly shows how the proposed method out- [9] M. L. Kerfi and D. Ziou, “Image retrieval based on feature weighing
performs the other compared methods even in terms of conver- and relevance feedback,” in Proc. IEEE Int. Conf. Image Processing
gence. As a matter of fact, the best compared method reaches (ICIP2004), 2004, vol. 1, pp. 689–692.
[10] A. Grigorova, F. G. B. De Natale, C. Dagli, and T. S. Huang, “Content-
its peak performance at convergence after ten iterations, while based image retrieval by feature adaptation and relevance feedback,”
Evolutionary PSO-RF reaches the same result after half of the IEEE Trans. Multimedia, vol. 9, no. 6, pp. 1183–1192, Oct. 2007.
iterations, then continuing its growth. A classical deterministic [11] G. Giacinto, F. Roli, and G. Fumera, “Content-based image retrieval
with adaptive query shifting,” in Machine Learning and Data Mining in
algorithm such as query shifting remains trapped in a local min- Pattern Recognition, P. Petra, Ed. Berlin, Germany: Springer-Verlag,
imum very soon, and the proposed method is able to achieve a 2001, vol. LNAI 2123, pp. 337–346.
similar result after just three to four iterations. Significant dif- [12] M. Koskela, J. Laaksonen, and E. Oja, “Use of image subsets in image
retrieval with self-organizing maps,” in Proc. Int. Conf. Image and
ferences are also reported with respect to other PSO implemen- Video Retrieval (CIVR), 2004, pp. 508–516.
tations. Typical PSO algorithms may require hundreds of itera- [13] G. Bordogna and G. Pasi, “A user-adaptive neural network supporting
tions to converge, while the proposed method converges in a few a rule-based relevance feedback,” Fuzzy Sets Syst., vol. 82, no. 9, pp.
201–211, 1996.
ones. This is mainly due to the interactive nature of our solution, [14] K.-H. Yap and K. Wu, “Fuzzy relevance feedback in content-based
where the convergence is guided by the repeated user feedbacks, image retrieval systems using radial basis function network,” in Proc.
which progressively add supervising knowledge to the problem. IEEE Int. Conf. Multimedia and Expo, 2005.
[15] D. Djordjevic and E. Izquierdo, “An object- and user-driven system for
semantic-based image annotation and retrieval,” IEEE Trans. Circuits
VII. CONCLUSIONS Syst. Video Technol., vol. 17, no. 3, pp. 313–323, Mar. 2007.
[16] Q. Tian, P. Hong, and T. S. Huang, “Update relevant image weights for
A new semi-supervised stochastic image retrieval method is content-based image retrieval using support vector machines,” in Proc.
presented in this paper. The image retrieval problem is formu- IEEE Int. Conf. Multimedia and Expo, 2000.
[17] X. S. Zhou and T. S. Huang, “Relevance feedback in image retrieval: A
lated as an optimization problem, and is solved by using a par- comprehensive review,” Multimedia Syst., vol. 8, pp. 536–544, 2003.
ticle swarm optimization. The user is in the loop and his interac- [18] M. J. Huiskes and M. S. Lew, “Performance evaluation of relevance
tion with the machine is used as an iterative supervision of the feedback methods,” in Proc. ACM Int. Conf. Content-Based Image and
Video Retrieval, 2008, pp. 239–248.
classification. The user’s feedback drives a feature re-weighting [19] N. Doulamis and A. Doulamis, “Evaluation of relevance feedback
process and the progress of the swarm, towards the goal of min- schemes in content-based in retrieval systems,” Signal Process.: Image
imizing a fitness function that takes into account the character- Commun., vol. 21, no. 4, pp. 334–357, 2006.
[20] E. O. Wilson, Sociobiology: The new synthesis. Cambridge, MA:
istics of the relevant and irrelevant images, as points of attrac- Belknap Press of Harvard Univ. Press, 1975.
tion and repulsion. Experimental results show the effectiveness [21] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in Proc.
of the proposed method. To enhance the retrieval performance, IEEE Conf. Neural Networks IV, Piscataway, NJ, 1995.
[22] R. C. Eberhart and Y. Shi, “Particle swarm optimization: Develop-
further studies are being conducted about the re-weighting and ments, applications and resources,” in Proc. Congr. Evolutionary Com-
fitness function. Different image databases and features spaces putation, 2001, vol. 1.
are being compared and analyzed to better customize the PSO [23] K. E. Parsopoulos and M. N. Vrahatis, “Recent approaches to global
optimization problems through particle swarm optimization,” Natural
algorithm. Further developments will try to insert into the loop Comput., vol. 1, no. 2–3, pp. 235–306, 2002.
explicit semantic knowledge (in the form of annotation) together [24] K. Chandramouli, “Particle swarm optimization and self-organizing
with an unsupervised clustering of the database. Furthermore, it maps based image classifier,” in Proc. 2nd Int. Workshop Semantic
Media Adaptation and Personalization, 2007, pp. 225–228.
is worth mentioning that standardized methodologies to allow [25] P. Yuan, C. Ji, Y. Zhang, and Y. Wang, “Optimal multicast routing in
measuring the user satisfaction based on subjective criteria are wireless ad hoc sensor networks,” in Proc. IEEE Int. Conf. Networking
still missing. Studies are being carried on to place the proposed Sensing and Control, 2004, vol. 1, pp. 367–371.
[26] D. Gies and Y. Rahmat-Samii, “Reconfigurable array design using par-
RF methodologies in a framework for human-oriented testing allel particle swarm optimization,” in Proc. IEEE Int. Symp. Antennas
and assessment. and Propagation, 2003, pp. 177–180.
[27] H. B. Liu, Y.-Y. Tang, J. Meng, and Y. Ji, “Neural networks learning
using VBEST model particle swarm optimization,” in Proc. 3rd Int.
REFERENCES Conf. Machine Learning and Cybernetics, 2004, pp. 3157–3159.
[1] Y. Rui, T. S. Huang, and S.-F. Chang, “Image retrieval: Current tech- [28] K. Chandramouli, T. Kliegr, J. Nemrava, V. Svatek, and E. Izquierdo,
niques, promising directions and open issues,” J. Vis. Commun. Image “Query refinement and user relevance feedback for contextualized
Represent., vol. 10, no. 4, pp. 39–62, 1999. image retrieval,” in Proc. VIE 08, 2008.

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.
BROILO AND DE NATALE: STOCHASTIC APPROACH TO IMAGE RETRIEVAL 277

[29] K. Chandramouli and E. Izquierdo, “Image classification using self [45] M. Clerc and J. Kennedy, “The particle swarm-explosion, stability, and
organising feature maps and particle swarm optimization,” in Proc. convergence in a multidimensional complex space,” IEEE Trans. Evol.
7th Int. Workshop Image Analysis for Multimedia Interactive Services Comput., vol. 6, no. 1, pp. 58–73, Feb. 2002.
(WIAMIS’06), 2006, pp. 313–316. [46] Y. Shi and R. Eberhart, “Parameter selection in particle swarm opti-
[30] M. Okayama, N. Oka, and K. Kameyama, “Relevance optimization in mization,” in Proc. 7th Annu. Conf. Evolutionary Programming, Mar.
image database using feature space preference mapping and particle 1998.
swarm optimization,” Neural Inf. Process., pp. 608–617, 2008. [47] I. C. Trelea, “The particle swarm optimization algorithm: Convergence
[31] K. Chandramouli and E. Izquierdo, “Image retrieval using particle analysis and parameter selection,” Inf. Process. Lett., pp. 317–325,
swarm optimization,” in Ser. Advances in Semantic Media Adaptation 2003.
and Personalization. Boca Raton, FL: CRC, 2008. [48] D. Bratton and J. Kennedy, “Defining a standard for particle swarm
[32] L. Wang, Y. Zhang, and J. Feng, “On the Euclidean distance of images,” optimization,” in Proc. IEEE Swarm Intelligence Symp., 2007.
IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 8, pp. 1334–1339,
Aug. 2005.
[33] J. Rocchio, “Relevance feedback in information retrieval,” in The
SMART Retrieval System: Experiments in Automatic Document Pro-
cessing, G. Salton, Ed. Englewood Cliffs, NJ: Prentice-Hall, 1971,
pp. 313–323.
[34] Y. Wu and A. Zhang, “A feature re-weighing approach for relevance Mattia Broilo (S’08) was born in 1982. He received
feedback in image retrieval,” in Proc. IEEE Int. Conf. Image Processing the B.S. and M.S. degrees in telecommunications en-
(ICIP2002), 2002, vol. 2, pp. 581–584. gineering from the University of Trento, Trento, Italy,
[35] R. Poli, J. Kennedy, and T. Blackwell, “Particle swarm optimization an in 2004 and 2007, respectively, where he is currently
overview,” Swarm Intell., pp. 33–57, 2007. pursuing the Ph.D. degree in telecommunications.
[36] M. G. Hinchey, R. Sterritt, and C. Rouff, “Swarms and swarm intelli- In 2005 and 2006, he joined different collabo-
gence,” Computer, vol. 40, no. 4, pp. 111–113, 2007. rations within the Department of Engineering and
[37] J. R. Robinson and Y. R. Samii, “Particle swarm optimization in Computer Science (DISI) of the University of Trento.
electromagnetics,” IEEE Trans. Antennas Propag., vol. 52, no. 3, pp. His interests include image and multimedia signal
771–778, 2004. processing, with particular attention to computer
[38] G. Griffin, A. Holub, and P. Perona, Caltech-256 Object Category vision applications for content-based image retrieval,
Dataset, California Inst. Technol., 2007, Tech. Rep. 7694. pattern recognition, and image processing.
[39] D. Tao, X. Tang, X. Li, and Y. Rui, “Direct kernel biased discrimi-
nant analysis: A new content-based image retrieval relevance feedback
algorithm,” IEEE Trans. Multimedia, vol. 8, no. 4, pp. 716–727, Apr.
2006. Francesco G. B. De Natale (M’96–SM’03) received
[40] S. C. Hoi, M. R. Lyu, and R. Jin, “A unified log-based relevance feed- the M.Sc. degree in electronic engineering in 1990
back scheme for image retrieval,” IEEE Trans. Knowl. Data Eng., vol. and the Ph.D. degree in telecommunications in 1994
18, no. 4, pp. 509–524, Apr. 2006. from the University of Genoa, Genoa, Italy.
[41] B. S. Manjunath, J.-R. Ohm, V. V. Vasudevan, and A. Yamada, “Color He is a Professor of telecommunications at the
and texture descriptors,” IEEE Trans. Circuits Syst. Video Technol., vol. University of Trento, Trento, Italy, is the Head of
11, no. 6, pp. 703–715, Jun. 2001. the Department of Information Engineering and
[42] M. Uysal and F. Yarman-Vural, “Selection of the best representative Computer Science (DISI), and is responsible for the
feature and membership assignment for content-based fuzzy image Multimedia Signal Processing and Understanding
database,” in Proc. Int. Conf. Image and Video Retrieval, Lecture Lab. His research interests include multimedia
Notes in Computer Science, 2003, vol. 2728, pp. 138–147. signal processing, analysis, and transmission, with
[43] T. Deselaers, D. Keysers, and H. Ney, “Features for image retrieval: particular attention to image and video data.
An experimental comparison,” Inf. Retriev., vol. 11, no. 2, pp. 77–107, Dr. De Natale was General Co-Chair of the Packet Video Workshop in 2000
2008. and Technical Program Co-Chair of the IEEE International Conference on
[44] S. Aksoy and R. M. Haralick, “Feature normalization and likelihood- Image Processing (ICIP) in 2005 and the IEEE International Conference on
based similarity measures for image retrieval,” Pattern Recognit. Lett., Multimedia Services Access Networks (MSAN, now Mobimedia) in 2005. In
vol. 22, no. 5, pp. 563–582, 2001. 1998, he was co-recipient of the IEEE Chester-Sall Best Paper Award.

Authorized licensed use limited to: Praveen V. Downloaded on July 20,2010 at 04:43:42 UTC from IEEE Xplore. Restrictions apply.