A Relevant Visual Feature Selection Approach for Image Retrieval

Olfa Allani

1,2

, Nedra Mellouli

, Hajer Baazaoui Zghal

, Herman Akdag

and Henda Ben Ghzala

LIASD, Universit

e Paris 8, Saint-Denis, France

LABORATOIRE RIADI, ENSI Campus Universitaire de la Manouba, Tunis, Tunisia

LIASD,EA 4383, Universit

e Paris 8 Saint-Denis, IUT de Montreuil, Montreuil, France

Keywords:

Visual Feature, Ontologies, Semantic Content, Selection.

Abstract:

Content-Based Image Retrieval approaches have been marked by the semantic gap (inconsistency) between

the perception of the user and the visual description of the image. This inconsistency is often linked to the

use of predeﬁned visual features randomly selected and applied whatever the application domain. In this

paper we propose an approach that adapts the selection of visual features to semantic content ensuring the

coherence between them. We ﬁrst design visual and semantic descriptive ontologies. These ontologies are

then explored by association rules aiming to link semantic descriptor (a concept) to a set of visual features.

The obtained feature collections are selected according to the annotated query images. Different strategies

have been experimented and their results have shown an improvement of the retrieval task based on relevant

feature selections.

1 INTRODUCTION

The development of efﬁcient image retrieval ap-

proaches is a very active research area. Content

based image retrieval approaches (CBIR) are partic-

ularly very popular because they are automatic and

do not rely on the users to perform the retrieval pro-

cess (Liu et al., 2007). Instead, they use internal de-

scription consisting of visual feature vectors to index

images. Visual feature vectors that index the image

database are numeric information extracted from im-

age pixels using image processing and analysis tech-

niques. In literature the problem of image indexation

has been addressed using two different visual feature

categories which are common visual features and ﬁeld

speciﬁc visual features (Mussarat et al., 2013).

Common visual features (low-level) can be color,

texture or shape features. They can be extracted from

a whole image (global approach) or from speciﬁc

regions (local approach) using clustering techniques

(Lavenier, 2001).

Field speciﬁc features are generally motivated by

a particular application context. For instance, to ad-

dress the face recognition problem, the Viola and

Jones detection framework has been proposed (Viola

and Jones, 2004). The work introduced in (Sarfraz

and Hellwich, 2008) propose a robust procedure for

face recognition based on a feature taking into ac-

count facial appearance shape and illumination.

Several low-level and domain speciﬁc features

have been proposed in the literature. However, the

accuracy of retrieval results remain far from users ex-

pectations (Smeulders et al., 2000). On one hand,

the adoption of sophisticated features can improve re-

sults, but highlights the high algorithms’ complexity,

the computation time and the feature vectors’ dimen-

sion problems (Vinukonda, 2011). On the other hand,

applying a large number of features disperses the re-

trieval focus and makes the process time consuming.

Our idea is to integrate a feature selection mech-

anism in the image retrieval process, aiming to guide

the selection of features to be used to index query im-

ages. The proposed selection allows retrieving im-

ages based on suitable features to the current query

image instead of applying random sets of visual fea-

tures. Our contribution consists in building feature

collections in line with speciﬁc semantic content, and

providing a dynamic selection mechanism of visual

feature sets. The selection mechanism that we pro-

pose is based on the user query annotation which is

considered by our approach as a relevant source of

semantic content description.

The remainder of this paper is organized in 3 sec-

tions. In section 2, we review some common CBIR

implementations. After discussing the study of re-

lated works, we detail our approach in section 3. The

implementations of the proposed approach as well as

the experimental results are introduced in section 4.

Finally, discussion and results analysis as well as fu-

ture works are presented in conclusion.

377

Allani O., Mellouli N., Baazaoui Zghal H., Akdag H. and Ghzala H..

A Relevant Visual Feature Selection Approach for Image Retrieval.

DOI: 10.5220/0005306303770384

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 377-384

ISBN: 978-989-758-090-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

2 OVERVIEW AND

MOTIVATIONS

In this section we present some related (but not ex-

haustive) works on image retrieval approaches using

low-level visual features. Then, we discuss the classi-

cal approaches problems and propose our motivations

to dynamic feature selection giving a query image.

2.1 Visual Features for CBIR

Approaches

Several low-level visual features, related to color, tex-

ture and shape; have been proposed in literature and

widely used in retrieval approaches. We review here

some common CBIR approaches.

The existing implementations of CBIR ap-

proaches are mainly based on different feature cate-

gories. The ﬁrst CBIR systems rely on one feature.

Since color is an intuitive image content descriptor

and simple to extract. It was used as a basis for sev-

eral implementations.

In (Pass et al., 1996), an approach based on his-

togram with a spatial dimension has been proposed.

The experiments have been done on a large diversiﬁed

image database and have shown the improvement of

the retrieval results compared to histogram based ap-

proaches. This approach and several other approaches

rely on color histogram or some variations of it. How-

ever, these approaches store the extracted information

for each image which may require signiﬁcant space.

Thus in (Paschos et al., 2003), an approach based on

the chromaticity diagram has been proposed. This ap-

proach reduces the required space and maintains the

effectiveness of the results compared to classical ap-

proaches.

Color is a restrictive parameter especially when

dealing with speciﬁc ﬁelds such as face detection or

footprint recognition. In this ﬁelds, the detection of

exact shapes seems to be necessary. Thus, another

class of approaches based on shape features has been

proposed. CBIR systems have been used for a diverse

range of images, however, shape detection algorithms

have usually been designed for special issues. We fo-

cus here on works that implement shape features for

image retrieval purpose.

In (Lin et al., 2004), an efﬁcient and robust shape-

based image retrieval system is proposed. The Prompt

edge detection method is used to detect edge points.

Then the low-to-high sequence (LHS) shape repre-

sentation method is introduced. The results proved

the method robustness and effectiveness. But it is

worth noting that shape based approaches have higher

complexity. In (Hejazi and Ho, 2007), an image re-

trieval approach based on classical texture features,

such as orientation, directionality, and regularity has

been proposed. Their discriminant power has been

compared to the MPEG-7 texture feature through ex-

perimentation on aerial images.

Other CBIR systems combining different cate-

gories of features have been proposed. In (Jalab,

2011), a color layout and Gabor texture descriptor

based approach has been proposed. The Color Layout

Descriptor (CLD) represents the spatial distribution

of colors in an image. However, Gabor texture fea-

ture describes the texture distribution of similar col-

ors in an image. The integration of the two features

has signiﬁcantly improved the retrieval performance.

Retrieval results were compared with two other ap-

proaches ((Hiremath and Pujari, 2007) and (Haﬁane

and Zavidovique, 2008)) and proved to be more accu-

rate.

The approaches presented above are entirely

based on visual aspects. In fact, they focus mainly

on the problem of designing new features and ﬁnd-

ing possible combinations between existing features.

These proposals do not take into account the users

perception usually expressed in textual annotations.

However, despite the large amount of approaches

such as machine learning, relevance feedback and on-

tologies, the research area is still open and new ap-

proaches appear every day. These approaches are

promising, but the use of visual features can be in-

consistent in some context with semantic descriptors.

The correlation of joint features that should be applied

is always given in order to reduce computation time of

retrieval processes, analyze feature performance or to

ﬁt data set speciﬁcity.

Unfortunately, despite this large number of imple-

mentations, several problems persist, mainly how to

select an initial set of suitable features for a speciﬁc

semantic content. In literature, the main approaches

randomly establish a number of visual features to be

applied during retrieval (Deselaers et al., 2008), ig-

noring in someway speciﬁc semantic information re-

lated to the query image and that could be beneﬁcial

in establishing a coherent understanding of the real

human perception.

2.2 Motivations

In the context of image indexing for retrieval pur-

poses, it can clearly be seen that content based ap-

proaches, using different features and applied on the

same image dataset, perform differently (Jalab, 2011).

As a result, the research quality heavily depends on

the selected features.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

378

In addition, image retrieval systems usually estab-

lish previously a given set of visual features which

are used for any application domain. Thus, for two

queries, it is possible to obtain completely different

precision values (a high precision of retrieval results

for one and a low value for the other), although the

retrieval process relies on the same features. In this

case, semantic information deduced from image an-

notation is ignored and retrieval is performed stati-

cally without taking into account the image speciﬁc

semantic content.

Furthermore, the selection of a set of visual fea-

tures for a query image that is guided by its semantic

content can be considered as an interesting idea.

Our purpose is ﬁrst to provide a set of relevant

visual features for each query image, given its seman-

tic content. This enables to improved retrieval results

accuracy, to substantially reduce the feature vectors

number and to decrease the processing time. So, the

aims of this paper can be summarized in:

• Building visual features collection according to

semantic content. These collections allow asso-

ciating, for a concept, a set of suitable visual fea-

tures that should be applied on queries containing

this concept. Suitability is deduced from the dif-

ferent application domains reviewed in literature

works.

• Integrating relevant visual feature selection dur-

ing the retrieval phase thus allowing a dynamic

retrieval process based on query image semantic

content. The feature selection uses image anno-

tation to select visual feature collection and then

extract the most relevant features.

3 RELEVANT VISUAL

FEATURES’ SELECTION FOR

IMAGE RETRIEVAL

The proposed approach is part of a research line with

wider scope, in which a hybrid retrieval approach has

been deﬁned (Allani et al., 2014). Our idea is to use

both visual features and textual content in order to

perform a pattern-based retrieval. Our work has fo-

cused on structuring image dataset into a set of pat-

terns which are semantically and visually rich. Then

we retrieve results using similarity measure computed

between query similarity and the patterns.

Let’s consider a query image composed of the

”sky”, two ”divers”, ”ocean”, ”sand beach” and the

”clouds”. This image represents different objects and

so can be described using different visual features.

Whereas the ”sky”, the ”ocean” and the ”sand beach”

are characterized by their uniform texture, the ”diver”

and the ”clouds” are characterized by their speciﬁc

shape (shape of a person, shape of the clouds). More-

over, the image represents many meta-data character-

istics. We aim here to apply, during retrieval of sim-

ilar images, features suitable to the semantic content

and the meta-data of the image.

Shape features can be used to index images rep-

resenting shapes, same for texture. Also, meta-data

characteristics can be used to select appropriate vi-

sual features. For example a high resolution involves

a time consuming processing for feature extraction,

so features with high complexity (for example region

based shape features) should not be applied with such

images. As a result, for this speciﬁc image, using tex-

ture and shape features such as Edge Histogram and

canny edge detector which is a contour based shape

feature, can provide more relevant retrieval results be-

cause they are suitable to the query image.

The overall architecture of our approach is illus-

trated in Figure 1. The process begins with building

a set of visual features’ collections dedicated to spe-

ciﬁc concepts or meta-data characteristics. Building

process is performed given a set of annotated images

and a set of suitability rules deduced form literature.

Image dataset can be updated when new images are

added to it.

As depicted in Figure 1, our retrieval process,

based on relevant feature selection, is performed in

two phases: online and ofﬂine phases. In the next

paragraph the different steps of our retrieval approach

will be detailed.

The ﬁrst step is to specify the set of candidate vi-

sual features on which selection mechanism will be

performed. Visual feature vectors are computed on

the image dataset and stored (cf. Figure 1 Step (1)).

They are then clustered into regions (cf. Figure 1 Step

(3)).

Concepts are then extracted from image annota-

tions. A disambiguation step, based on WordNet

, is

performed in order to retrieve the good synset (sense)

that corresponds to each word (cf. Figure 1 Step (2)).

The set of concepts associated to all of the image

dataset is stored. Finally, a uniﬁcation step based on

WordNet and aiming to get common super-concepts

is performed. For example, two image annotations

containing the words ”Laguna Colorado” and ”Green

Lake” will be processed. The co-occurrences of the

two words are substituted by their lowest common an-

cestor which is ”lake”.

The previously described steps allow getting a vi-

sually and semantically indexed image dataset. Next

we deﬁne, given the semantic content of the dataset,

http://wordnet.princeton.edu/

ARelevantVisualFeatureSelectionApproachforImageRetrieval

379

Figure 1: Relevant visual feature selection based approach.

the visual feature collection. Image semantic content

(concepts) and meta-data are ﬁrst associated to ap-

propriate visual features. These associations known

as Application rules (cf. section 3.2) deﬁne the re-

latedness between semantic content or meta-data and

low-level visual features. They are deduced from the

best practices or benchmarks collected from literature

works on CBIR implementations. In order to build a

valid set of rules, reﬂecting high correlation between

semantic and visual aspects, we consider about 40 lit-

erature works covering several low-level visual fea-

tures and different image datasets.

Thereafter, based on the concepts used to describe

the image dataset, we propose a mechanism for build-

ing visual features’ collections that are adapted to se-

mantic content of query images. This approach analy-

ses concept simultaneous presence frequencies to out-

line relevant visual features for a given query image.

For each concept, related concepts (used with it in

an image annotation) are collected and their presence

frequencies for each image are analysed. Inconsistent

concepts with frequencies lower than a threshold α

are removed. In contrast consistent concepts are con-

sidered to build collections of visual features based on

application rules (cf. Figure 1 Step (4)). These collec-

tions will be used next to select relevant visual feature

collection suitable to query images.

During the online phase, a query image annotated

or not is introduced. If the query image annotated,

then concepts are extracted from annotation and dis-

ambiguation is performed (cf. Figure 1 Step (5)).

When visual feature collection are selected for all the

concepts, the number of visual features can be high.

Thus, a reasoning mechanism should be applied to

keep the most relevant features (cf.Figure 1 Step (6)).

Finally, a similarity measure is used to retrieve similar

images (cf. Figure 1 Step (8)).

3.1 Visual Features, Meta-data and

Image Content Ontologies

The proposed approach takes into account the seman-

tic content and the meta-data characteristics of the

query image ﬁle, to set up relevant low-level feature

collection. In order to organize the various informa-

tion, we need to formally represent knowledge within

each of the cited domains.

Ontologies are powerful knowledge representa-

tion tools that provide an explicit speciﬁcation of

knowledge in a structured and organized manner.

They allow enrichment and direct access to several

types of relationships between the different concepts

(Besbes and Baazaoui Zghal, 2014).

Thus, we propose to model three ontologies: image

semantic content, meta-data and low-level features.

First we build a semantic hierarchy integrating

concepts used to describe images textual content.

Each image in the database should be represented by

at least one concept of the image semantic content on-

tology.

Figure 2 illustrate the semantic content ontology

which is a uniﬁed characterization of images seman-

tic content. Each node of this hierarchy represents a

concept and each edge represents the relationship ”Is-

a”. An image can be described by a set of concepts

and their relationships.

Deﬁnition 1 (Image Content Ontology). Let O

the image content ontology. The ontology concepts

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

380

Figure 2: Image content ontology.

are noted:

C =

{

,··· , c

}

is the set of concepts represented by

the leaves of the ontology O

Non annotated images do not have rich semantic de-

scription, however, they rely on the surrounding tex-

tual information as meta-data. Meta-data provide use-

ful information on images namely the image author,

device, format, description, resolution etc.

We build a hierarchy to represent meta-data char-

acteristics as well as the possible values of each char-

acteristic. Image meta-data ontology is more or less

sophisticated hierarchy established to provide a uni-

ﬁed description for image characteristics in spite of

characteristics values heterogeneous nature. In the

proposed hierarchy, nodes represent image ﬁle char-

acteristics and associated values, and edges represent

the relationship ”Is-a”. Each image is represented

with many characteristics values of the meta-data on-

tology.

Deﬁnition 2 (Meta-data Ontology). Let O

be the

image meta-data ontology.

Image meta-data characteristics are noted:

M =

{

,··· , m

}

A characteristic m

,i ∈ 1···v, m

is associated to

membership values.

Let V

be the set of possible values associated to m

{

,··· , v

}

deﬁned on (M × D

∈M

) for k ∈

1···v

For example, an image has a ”high” resolution, a low

size, ”nature scene” as a title and ”JPEG” extension.

Each characteristic of the image ﬁle could be rep-

resented by a set of values (depending on the char-

acteristic domain). For example, if m

represents the

resolution, we associate to this characteristic the val-

ues ”low”, ”average” and ”high”.

In order to manage visual features studied in dif-

ferent literature works, we build an ontology for vi-

sual features (cf. Figure 3). This ontology has the

advantage of being extendible (a new feature or a new

use of an existing feature) and to provide an organized

description taking into account possible relationships

between features. Thus, features ontology is a uniﬁed

characterization of low-level features. it is a hierarchy

where the nodes represent features categories (shape,

color, color), application mode(global, local) and ap-

plication domains (speciﬁc, generic). Visual features

from different categories and with different use are

integrated in this hierarchy. The leaves represent vi-

sual features and the edges represent the relationship

”Is-a”.

Deﬁnition 3 (Visual Features Ontology). Let O

the visual features ontology.

We note F the set of available visual features:

F =

{

,··· , f

}

Figure 3: Visual features ontology.

3.2 Application Rules

The analysis of several literature works allowed us to

identify several feature application domains. For in-

stance, edge histogram is used in several works on

natural image datasets and provides efﬁcient retrieval

results. Thus, we make use of this observation and ex-

tract an application rule assuming that edge histogram

is associated to natural images. Following the above

assumption, we build a set of application rules for a

set of candidate visual features related to color, shape

and texture.

We use the previously deﬁned ontologie notations

of the concepts in rules reformulation. As a result,

each application rule is expressed according to a con-

cept from O

or from O

and a concept from O

However, application rules correspond usually to

general concepts such as nature, object or texture.

However, concepts in query image annotation are

very speciﬁc. Thus, using the image content ontology

ARelevantVisualFeatureSelectionApproachforImageRetrieval

381

as well as WordNet allows deducing implicit appli-

cation rules. For example, Building → Color-Layout

is a rule deduced from literature works. A query im-

age with a ”hospital” or a ”farm house”, which are

concepts related to building, will be associated to the

feature Color Layout.

Deﬁnition 4 (Application Rules). Let I =

{

,··· , I

}

is the set of images in the database where N = ]I is

the total number of images.

We note R

a feature application rule:

{

}

l∈1···p

→ c

, where 1 ≤ p ≤ k and such that

]

{

∈ D

}

0 .

3.3 Visual Feature Collections: Building

and Selection

As presented above, our goal is to ﬁnd out relevant

features given the semantic content and to decide

which feature collection should to be applied together.

To deal with the second problem, we introduce the

collections of features idea. A collection of features

is a set of low-level features suitable to a speciﬁc con-

cept or characteristic. In order to create the feature

collection, we adopt an inductive reasoning (Akdag

et al., 2000) in which the premises seek to supply evi-

dence. In our case, the premises are the concepts used

in annotation. As previously precised, concepts used

to annotate the image dataset are uniﬁed and stored.

For each image a ﬁnal set of concepts is identiﬁed.

For each concept, we extract images where this con-

cept has been used. Then, we compute appearance

frequencies of other concepts in these images. This

allows to deduce which concepts are frequently used

with the concerned concept.

Deﬁnition 5 (Feature Collection Building). Let w

the ﬁnal set of concepts associated to the image I

{

···c

}

are the concepts in w

For the concept c

let I

be the set of images where c

exists. m is the cardinality of I

For the set I

, w



···c

l p



are the concepts si-

multaneously used with c

We compute f

··· f

l p

which are the numbers of oc-

currences of c

···c

l p

simultaneously with c

from the

images in I

devided by m.

Concepts c

with f

≥ α are retained.

We also apply treatments on image meta-data in order

to obtain a uniﬁed description of meta-data. Finally,

on the whole image collection we compute, for every

speciﬁc concept, other concepts presence frequencies

in order to deduce the possible feature collections.

The obtained set of concepts associated to the con-

cerned concept allows to deduce the feature collec-

tion. A relationship between the concepts and the im-

age content ontology is deduced based on WordNet

relationships. Then, using these relationships we de-

duce for each concept an application rule. As a result

we obtain a set of features that we associate to the

concerned concept. It is worth noting that each col-

lection is characterized by a priority factor consisting

of the ratio of the images where the concerned con-

cept appeared and the total number of images in the

database. Moreover, image dataset update leads nec-

essarily to changing in the concepts. In this case, col-

lections update is also needed.

When a query image is introduced, image anno-

tation is processed in order to extract concepts and

use them to select relevant features to apply during

the retrieval process. To each concept, a collection of

features is proposed. In order to get the ﬁnal features

collection, a reasoning step taking into account prior-

ity of the different collections is performed. Finally,

given the selected features, query image is indexed

and similarity is computed (cf. Figure 4 ).

Figure 4: Visual feature selection for image retrieval.

4 EXPERIMENTATION AND

RESULTS ANALYSIS

Our proposal has been implemented and evaluated in

order to show its interest in image retrieval. From

more technical point of view, the frameworks that

have been employed are Java as programming lan-

guage. The source codes of visual features imple-

mentations are provided by the jFeatureLib

and the

LIRe

libraries.

https://code.google.com/p/jfeaturelib/

https://code.google.com/p/lire/

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

382

Different strategies have been deﬁned to allow the

evaluation of the improvement degree provided by

our approach. Query images are classiﬁed given their

content into 6 classes. On each class of query images

we perform 7 retrieval strategies given the feature(s)

categories. The goal is to get the higher precision and

recall values for a strategy conform to an image class.

For instance, the best value of precision and recall ob-

tained for textural images correspond to texture based

retrieval.

To evaluate our approach, we used the Image Clef

2013 dataset

. There are 117 query images (with the

qrels ﬁle) in different sizes (average image dimen-

sions: 480*360 pixels) and representing various ob-

jects and themes.

4.1 Experimental Setup

For measuring the image retrieval effectiveness we

used as evaluation metrics:

• The exact precision measure (P@10)

• The recall measure

For the experiments we apply the 7 following re-

trieval strategies: SBIR (shape strategy with canny

edge detector); TBIR (texture strategy with edge his-

togram and the Gabor ﬁlter); CBIR (color strategy

with color layout and scalable color descriptors); TS-

BIR (texture and shape strategy with edge histogram

and canny edge detector); CSBIR (color and shape

strategy with color layout and canny edge); CTBIR

(color and texture strategy with color layout and edge

histogram); SCTBIR (shape, color and texture strat-

egy with canny edge detector, color layout and edge

histogram). These strategies are applied each time on

the following 6 classes of query images in order to

evaluate the impact of our feature selection approach

given the image semantic content: S-Class ( images

with object shapes) ; T-Class (images with textures);

TS-Class (textural and shape images); CS-Class(color

and shape images); CT-Class (color and textural im-

ages); TSC-Class (texture shape and color combined

images).

4.2 Evaluation Results

Figure 5 shows the results in term of precision for

the top 10 retrieved images according to the proposed

strategies. The obtained precision result for the strat-

egy texture based image retrieval (TBIR) is clearly

higher than other strategies when applied on texture

images class (CT). The same observation could be

http://imageclef.org/SIAPRdata

Figure 5: Precision.

noted when the strategy applied is associated to the

image class.

Figure 6 shows the results in term of recall. The

recall values are higher for the shape and texture

based image retrieval (STBIR) strategy when applied

on the shape and texture class. This allows to de-

duce the impact of relevant feature selection on the

retrieval.

Figure 6: Recall.

To complete these results, we computed the im-

provements for image class of the adapted strategy in

front of the average precision and recall values of the

other strategies.

Figure 7 and ﬁgure 8illustrate retrieval results us-

ing the CTBIR strategy and the SCTBIR strategy.

5 CONCLUSION

In this paper, an image retrieval approach has been de-

ﬁned. It relies on feature set collections building and

relevant feature selection mechanisms. These mech-

anisms allow a dynamic low-level feature selection

guided by the query image semantic content. In this

work, we have conducted an elementary experimen-

tal study where we are focused on the improvement

given by our approach. It is worth to be noted that

the preliminary results obtained with a targeted se-

lected features are more relevant than those obtained

ARelevantVisualFeatureSelectionApproachforImageRetrieval

383

Figure 7: Retrieval results using CTBIR strategy.

Figure 8: Retrieval results using SCTBIR strategy.

with features selected randomly. An advanced imple-

mentation and experiments are in progress to evaluate

the proposed approach: ontologies, rules and feature

building.

REFERENCES

Akdag, H., Mellouli, N., and Borgi, A. (2000). A sym-

bolic approach of linguistic modiﬁers. In Informa-

tion Processing and Management of Uncertainty in

Knowledge-Based Systems (IPMU).

Allani, O., Baazaoui-Zghal, H., Mellouli, N., Ben-Ghezala,

H., and Akdag, H. (2014). A pattern-based system

for image retrieval. In International Conference on

Knowledge Engineering and Ontology Development.

Besbes, G. and Baazaoui Zghal, H. (2014). Modular ontolo-

gies and cbr-based hybrid system for web information

retrieval. Journal of Multimedia Tools and Applica-

tions.

Deselaers, T., Keysers, D., and Ney, H. (2008). Features for

image retrieval: An experimental comparison. Infor-

mation Retrieval, 11.

Haﬁane, A. and Zavidovique, B. (2008). Local relational

string and mutual matching for image retrieval. Infor-

mation Processing & Management, 44.

Hejazi, M. R. and Ho, Y.-S. (2007). An efﬁcient approach

to texture-based image retrieval. International Journal

of Imaging Systems and Technology, 17.

Hiremath, P. S. and Pujari, J. (2007). P. s. hiremath and

jagadeesh pujari content based image retrieval based

on color, texture and shape features using image and

its complement.

Jalab, H. A. (2011). Image retrieval system based on color

layout descriptor and gabor ﬁlters. In Open Systems

(ICOS), 2011 IEEE Conference on. IEEE.

Lavenier, D. (2001). Paralllisation de lalgorithme du k-

means sur un systme reconﬁgurable application aux

images hyper-spectrales. In Traitement du Signal Vol-

ume 18.

Lin, H.-J., Kao, Y.-T., Yen, S.-H., and Wang, C.-J. (2004).

A study of shape-based image retrieval. In Distributed

Computing Systems Workshops, 2004. Proceedings.

24th International Conference on. IEEE.

Liu, Y., Zhang, D., Lu, G., and Ma, W.-Y. (2007). A sur-

vey of content-based image retrieval with high-level

semantics. Pattern Recogn., 40.

Mussarat, Y., Sharif, M., and Mohsin, S. (2013). Use of

low level features for content based image retrieval:

Survey. Research Journal of Recent Sciences, 2277.

Paschos, G., Radev, I., and Prabakar, N. (2003). Image

content-based retrieval using chromaticity moments.

Knowledge and Data Engineering, IEEE Transactions

on, 15.

Pass, G., Zabih, R., and Miller, J. (1996). Comparing im-

ages using color coherence vectors. In Aigrain, P.,

Hall, W., Little, T. D. C., and Jr., V. M. B., editors,

ACM Multimedia.

Sarfraz, M. S. and Hellwich, O. (2008). Head pose esti-

mation in face recognition across pose scenarios. In

International Conference on Computer Vision Theory

and Applications. INSTICC - Institute for Systems

and Technologies of Information, Control and Com-

munication.

Smeulders, A. W. M., Worring, M., Santini, S., Gupta, A.,

and Jain, R. (2000). Content-based image retrieval at

the end of the early years. IEEE Trans. Pattern Anal.

Mach. Intell., 22.

Vinukonda, P. (2011). A Study of the Scale-Invariant Fea-

ture Transform on a Parallel Pipeline. PhD thesis,

Louisiana State University.

Viola, P. and Jones, M. (2004). Robust real-time face detec-

tion. International Journal of Computer Vision, 57.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

384