Multimedia Indexing and Retrieval: Optimized Combination of

Low-level and High-level Features

Mohamed Hamroun

, Henri Nicolas

and Benoit Crespin

Department of Computer Science, Limoges University, XLIM Laboratory, Bordeaux, France

Department of Computer Science, Bordeaux University, LABRI Laboratory, Bordeaux, France

Keywords: Multimedia Indexing, Muti-Level Concepts, Multimedia Retrieval, Machine Learning.

Abstract: Nowadays, the number of theoretical studies that deal with classification and machine learning from a general

point of view, without focusing on a particular application, remains very low. Although they address real

problems such as the combination of visual (low-level) and semantic (high-level) descriptors, these studies

do not provide a general approach that gives satisfying results in all cases. However, the implementation of a

general approach will not go without asking the following questions: (i) How to model the combination of the

information produced by both low-level and high-level features? (ii) How to assess the robustness of a given

method on different applications ? We try in this study to address these questions that remain open-ended and

challenging. We proposes a new semantic video search engine called “SIRI”. It combines 3 subsystems based

on the optimized combination of low-level and high-level features to improve the accuracy of data retrieval.

Performance analysis shows that our SIRI system can raise the average accuracy metrics from 92% to 100%

for the Beach category, and from 91% to 100% for the Mountain category over the ISE system using Corel

dataset. Moreover, SIRI improves the average accuracy by 99% compared to 95% for the ISE. In fact, our

system improves indexing for different concepts compared to both VINAS and VISEN systems. For example,

the value of the traffic concept rises from 0.3 to 0.5 with SIRI. It is positively reflected on the search result

using the TRECVID 2015 dataset, and increases the average accuracy by 98.41% compared to 85% for

VINAS and 88% for VISEN.

1 INTRODUCTION

In the last decades, video occupies a remarkable place

in the information retrieval community. This can be

explained by the availability of digitized video data,

which is becoming more and more voluminous. This

constant growth in the mass of video data causes

problems for indexation, retrieval, and navigation.

Finding a desirable video in a huge video archive

is a very difficult task. In fact, users waste a lot of

time trying to fulfill their needs. Research in the field

of optimizing video data search and visualization

tools will become a major area of research in the

future. Indeed, the use of the video format is more and

more common. Also, the need to have powerful tools

to exploit this amount of data becomes important.

Currently, the majority of different approaches is

trying to improve these tools by relying on visual

content and semantic content (or low-level and high-

level descriptors). Although some works are efficient,

they could be improved as for example only a few of

them are based on multi-level fusion-based indexing,

which implies a combination between low-level and

high-level descriptors.

This article deals with the fusion of low-level

descriptors and high-level descriptors. We have

suggested a “SIRI” indexing approach to facilitate the

retrieval and the data usage. The innovative aspect of

the proposed system is the combination of low-level

and high-level descriptors. The main goal is to

guarantee both the accuracy of the results and of the

semantics. More precisely, we combine the low-level

descriptors we have proposed in other works, PMC

and PMGA (Hamroun et al), with high-level ones.

Moreover, we merge the descriptors that we have

already proven to be effective by integrating the

methods related to the ISE, VISEN and VINAS

systems.

2 RELATED WORKS

Information retrieval is a set of operations used to

retrieve a request through a user interface. First, the

194

Hamroun, M., Nicolas, H. and Crespin, B.

Multimedia Indexing and Retrieval: Optimized Combination of Low-level and High-level Features.

DOI: 10.5220/0011063000003179

In Proceedings of the 24th International Conference on Enterprise Information Systems (ICEIS 2022) - Volume 1, pages 194-202

ISBN: 978-989-758-569-2; ISSN: 2184-4992

user must formulate a query. This operation is evident

for text, but it is difficult for images and even more

difficult for videos. The query can be expressed by

different media. It can be an image, a video, a sound

or a sketch. The user can also include keywords. The

definition of queries is considered a difficult problem

with large-scale video databases. In the following we

present several existing forms of queries: image,

navigation, textual and conceptual queries.

2.1 Image Queries

This approach consists in introducing an example

image which will be used to find similar content. For

this kind of query, the indexing phase consists in

extracting a number of image descriptors. Color

features are among the most important recognition

aspects, as color remains an unchanging parameter

that is not altered by changes in the image orientation,

size or placement (Kundu et al). Content-Based

Image Retrieval (CBIR) systems use conventional

color features such as dominant color descriptor

(DCD) (Wang et al), color coherence vector (CCV)

(Pass et al), color histogram (Singha et al) and color

auto-correlogram (Chun et al). DCD is about

quantifying the space occupied by the color feature of

an image by placing its pixels into a measurable

number of partitions and calculating the means and

ratio of this placement. CCV partitions histogram

bins into coherent or incoherent types; the results of

this method are more precise since they not only

emanate from color histogram classification but also

from spatial classification. The accuracy of these

results is more visible when it comes to images that

contain rather homogeneous colors (Pass et al).

2.2 Navigation Queries

In the last decade several systems have been proposed

favoring a search in videos collections by navigation

in a tree of concepts. A recent review is given in

(Schoemann et al). In (Ben Halima et al), the authors

propose a video retrieval system by viewing and

navigating in concepts. The proposed navigation

module is based on a semantic classification. The

strength of this system is the ability to integrate a

personalized navigation module. The proposed

module is based on a neuronal metaphor. For

example, (Villa et al). present a system called "Facet

Browser", where searching for videos is based on the

simultaneous exploration of several search "facets".

In (Ben Halima et al) (Hamroun et al), a suggestion

tool is made available to assist the search of videos.

Two types of suggestions are considered, textual or

visual. The effectiveness of this system is proven by

analyzing the log files and with user studies.

(Schoeffmann et al). present an "instant video

navigation" system that segments the videos starting

from sequences visualized by the users. This tool

offers two different views for searching videos,

parallel or tree-based.

2.3 Textual Queries

Textual queries remain limited and are generally

associated with families of specific visual documents,

such as TV news. The most promising way in this

context regarding videos consist in transcribing the

soundtrack to determine its subject (Lefèvre et al),

rather than exploiting the visual content.

2.4 Conceptual Queries

Conceptual queries are the subject of many research

works. For example, the INFORMEDIA approach

uses a limited set of high-level concepts to filter the

results of textual queries. This system also creates

groups of key-images, using the results of speech

recognition to trace the collections of key-images at

the re-associated geographic place on a map, and

combines this with other visualizations to give the

user an arrangement of the query result context. The

suggested method in (Worring et al) is based on a

process of semantic indexing. This system uses a

large semantic lexicon divided in categories and

threads to support the interaction. It defines a space

of visual similarity, a space of semantic similarity, a

semantic thread space and browsers to exploit these

spaces. We can also mention the VERGE approach

(Stefanos et al), which supports conceptual and high

level visual research. This tool combines indexing,

analysis and recovery techniques of diverse

modalities (textual, visual and conceptual).

Many research works study the audiovisual data

search domains on large databases and suggest tools

based on diverse research forms. The limit of these

techniques often lies at the level of user interactions.

Enriching the search system with the past behavior of

the users potentially offers more pertinent results. In

the literature, only a few works focus on this aspect.

The approach suggested in (Faudemay et al) is the

first to put the user at the center of the search system.

This idea represents a real improvement perspective

to extend existing methods.

For the indexing part, most of the works presented

above are based on either a low level or a high-level

indexation and only a few are based on multi-level

indexing techniques. Thus, a very challenging

Multimedia Indexing and Retrieval: Optimized Combination of Low-level and High-level Features

195

Figure 1: Conceptual architecture of the SIRI Framework.

perspective is to enrich indexing techniques with a

multilevel-based indexing model. Information fusion

is an area which is undergoing significant

development, in particular for multimedia indexing

and retrieval. In this field, the data source is multiplied

with different characteristics (image, audio, text and

movement, etc), but as each data source is generally

insufficient, it is important to combine several

descriptors in order to gain better knowledge. Several

classes of combination methods have been proposed in

the literature. In (Kuncheva et al), Kuncheva makes the

difference between fusion and selection. Fusion

involves combining all the outputs, while selection

involves choosing the “best” from a set of possible

classifiers to identify the unknown form. (Duin et al)

distinguishes 2 types of combination methods in

fusion: (l) combination of different classifiers and (2)

weak classifiers. The latter has the same structure, but

is trained on different data. Other classes are proposed

by Xu, which distinguishes combination methods only

by the output type of the classifiers (class, measure)

presented at the input of the combination. Jain (Duin et

al) constructed a dichotomy according to two criteria:

the type of outputs from the classifiers and their

learning capacity. This last criterion is also used by

(Kuncheva et al) to separate different fusion methods.

Larning methods allow to find and adapt the para-

meters to be used in the combination according to the

database of available examples. The methods running

without learning simply use the outputs of the classi-

fiers without integrating in advance other information

on the performance of each of the classifiers.

3 SIRI: TOWARDS A

MULTI-LEVEL FRAMEWORK

Figure 1 shows the overall conceptual architecture of

our SIRI framework. It includes (i) an indexing phase,

which introduces a new merging approach detailed in

the following section, (ii) a phase composed of three

types of searches: textual, navigational and

exemplary image. The textual search is similar to the

VISEN system, while the navigation search is

inspired by our VINAS tool, and the exemplary image

is inspired by our ISE system.

Figure 2: Our multi-level algorithm.

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

196

Figure 2 includes an algorithm that provides an

overview of our multilevel fusion-based indexing

method. The following sections detail each of the

steps of this algorithm.

3.1 Step 1: Matching Concept /

Descriptor

Each Large scale Conceptual Ontology for

Multimedia (LSCOM) is better represented by one or

more low-level descriptors. For example, color

descriptors may be more determinant for some

concepts, such as sky, snow, water landscape or

vegetation, but less determinant for studios and

meetings. For this purpose, we propose to weight

each descriptor at a low level according to its degree

of discrimination in relation to the concept. Based on

our previous indexing and semantic studies

(LAMIRA and VISEN), for each concept we can

select the most relevant video shots and then model

this relationship in the form of a tree structure (figure

3) where the concept C

represents the root of tree A,

the vertices represent the set of videos (V

, V

and

V50 in our example), and the leaves of A represent

the set of shots (P

, P

in our example).

This tree is limited to only 3 levels and has vertices

that can be ordered according to their distance from

, withthe distance D(C

) defined as the weight of

the concept C

in shot P

Figure 3: Initial representation of the concept.

After selecting the designs or the weight of C

0.7 (value chosen based on an experimental research

study), we selected the low-level descriptors that

better represent each concept. These descriptors can

be of several types, such as: Hsv Histogram, Gabor

Texture, Camera Motion and Edge Histogram.

Moreover, the choice of a descriptor for a concept is

determined by its distance from that concept. This

value is chosen according to the SVMs classification

on the TRECVID 2015 data. For example, in figure

4, Edge Histogram, which is the most relevant

descriptor for all concepts, is characterized by

outlines, particularly for the concepts 'Car', 'Maps',

Mountain, Sports, Waterscape, unlike Camera

Motion, which influences only a few concepts. In

fact, it can have the same performance as that of the

rest of the descriptors for the 'Walking' and 'Running'

concepts.

Figure 4: Performance of the SVM classification per

concept for the four descriptors.

Table 1 shows the result of the concept/descriptor

relationship where each concept is defined by a set of

low-level descriptors.

Table 1: List of concepts according to the descriptors used.

3.2 Creating a Visual Dictionary

The first step consists in the extraction of visual

descriptors, such as colors, textures, etc, from each

corpus shot. It is therefore necessary to submit a

summary of the descriptors in a form more usable by

our system. For this purpose, a visual dictionary is

computed using the K-means clustering algorithm,

which has been selected due to its simplicity of

implementation and its speed of execution, which

enables us to determine a predefined number of centers

that best represent the set of descriptors. In our

experiments we used 130 centers (also called visual

keywords) related to semantic concepts. Figure 5

illustrates the principle of the visual dictionary

construction.

Multimedia Indexing and Retrieval: Optimized Combination of Low-level and High-level Features

197

Figure 5: Construction of the Visual Dictionary.

After creating the visual dictionary, which

contains all the descriptor vectors for each shot, we

connected the closest descriptors (i.e. the descriptors

that belong to the same class). Figure 6 shows an

example of these relationships. Moreover, the

descriptors represented in this example are CCV

Histo

, PMC

, MC

, Histo

, etc., which are examples

of the descriptors for P

, P

and P

. Therefore, the

closest descriptors in the visual dictionary, such as the

histograms 1 and 19, are related. Then, a simple

Euclidean distance was computed for each pair of

descriptors. Through this relationship and since these

descriptors are determinant for P

and P

, a

relationship is implicitly created between plans 1 and

19.

Figure 6: Extraction of the visual dictionary from concept

3.3 Selection of Descriptors for Each

Concept

We are now able to determine the descriptors that best

represent a concept for each shot. Therefore, in figure

7 we added the leaves CCV

, Histo

, PMC

, MC

and Histo

to concept C

present in plans 1, 19, 25, 50

and 51, building a tree with 4 levels instead of 3.

Figure 7: Selection of descriptors for each concept.

3.4

Matching between Descriptors

A matching is carried out between C

descriptors

(such as Histo

, DCD

, CCV

1, Histo

, etc.) and

the dictionary based on the K-means classification

algorithm. As shown in figure 8, the content of the

dictionary is divided into 5 classes (number of C1

descriptors) with a threshold s <=0.3 chosen

according to the experiment.

Figure 8: Matching between shots in the K classes.

The K-means algorithm helps group

homogeneous shots into K classes. In this method,

each class C

is represented by its gravity center,

which is the average of the descriptor vectors of the

shots belonging to this class.

Figure 9: Classification with K-means.

Figure 9 shows how the K-means algorithm

operates. We can distinguish 5 classes of shots of

which C

, C

and C

are three classes that represent

similar shots. The gravity centers of these three

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

198

classes are presented in the form of three vectors

reflecting the mean of nearest neighboring

descriptors. The other classes C

and C

group

together isolated shots which are not similar. In fact,

the objective of this method is to select the class

centers so as to reduce the sum of the intra-class

inertia within all the classes. The intra-class inertia of

class C

(i<n, where n is the number of classes),

denoted Inertia (C

), is the mean distance between the

vectors of the class and its center of gravity g

. Let P

be a shot, such that j < m where m is the total number

of shots in the set:

𝐼𝑛𝑒𝑟𝑡𝑖𝑎



𝐶





𝐶

𝐼𝑛𝑒𝑟𝑡𝑖𝑎



𝐶









∈

𝐼𝑛𝑒𝑟𝑡𝑖𝑎



𝐶







𝐶



𝑑𝑔



,𝐼









∈



𝑔





𝐶



𝐼







∈



Figure 10 shows the tree resulting from the

application of the descriptor classification, where P

, and P

are the new shots indexed by C

Figure 10: The result of the new classification.

3.5

Weighting

3.5.1 Concept Weighting in a Shot

Furthermore, we propose a new approach to calculate

the conceptual weighting by adding the concept of the

relationship between low-level and high-level

descriptors.

For our example, this value obtained is :

3.5.2 Weighting a Concept in a Video

𝑊𝐶𝑖,𝑉𝑗 

∑/



with N the total number of shotes in V

4 EXPERIMENTATION AND

RESULTS

4.1 Datasets

For the TRECVID 2015 Semantic Indexing task there

are two data sets provided by the National Institute of

Standards and Technology (NIST): one for testing

and one for training. The training dataset named

IACC.1.tv10.training contains 3200 internet archive

videos (50GB, 200 hrs) while the test dataset named

IACC.1.A contains approximately 8000 internet

archive videos (50GB, 200 hrs), annotated with 130

semantic concepts.

4.2 The User’s Interface

The interface shown in figure 11 represents the main

menu interface of our image research system. Part (A)

helps the user to search per keywords in English,

French or Arabic with a relevance feedback

mechanism. Then, in part (B), the user can navigate

through a hierarchical structure (context/ concept/

image) to get the desired image. Finally, part (C)

enables the user to search through an image query

from an image database containing 10000 examples.

Figure 11: Main menu of the semantic research system.

4.3 Step 1: Our System versus Low

Level Systems

The average precision for all the image categories of

the Corel-1k dataset is reported in figure 12. To show

the utility of our CBIR scheme, the results of nine

other CBIR systems are also reported in this table.

Since the average precision of our results is 99%, our

CBIR scheme has the highest accuracy among the

other state-of-the-art CBIR systems. In fact, our

proposed CBIR system outperforms that of (Jhanwar

Multimedia Indexing and Retrieval: Optimized Combination of Low-level and High-level Features

199

Figure 12: Comparison of the average precision of the previous methods and proposed method.

Figure 13: Comparison of the average precision of the (Elleuch et al),( Waseda et al),( Xiao et al), SVI_LAMIRA (Hamroun

et al) tools and proposed SVI_SIRI.

et al.), (Lin et al.), (ElAlami et al), (Murala et al.),

(Yildizer et al.), (Kundu et al), (Dubey et al), (Aun et

al), (Dubey et al), (Ashraf et al), (Walia et al.), (Zeng

et al.), (Zhou et al.), (Shrivastava et al.) and (Xiao et

al.).

4.4 Step 2: Our System versus

Detection of Concepts Systems

Figure 13 represents the weighting level of the 30

concepts. These weighting levels are variable

according to the systems. We have compared our

SVI_LAMIRA (Hamroun et al) system with (Elleuch

et al),( Waseda et al) and (Dubey et al).

For some concepts such as Bicycling, Singing,

Telephones and Car Racing, SVI_LAMIRA obtained

the best results. Thus, the use of a semantic

interpretation process improves the detection of

semantic concepts by using our DCM method

(Hamroun et al)

4.5 STEP3: Our System versus Height

Level Systems

In the 2nd step of the assessment, we compare our

system with the most pertinent semantic retrieval

systems. The following figure 14 shows some query

outcomes.

Based on the histogram shown on figure 14, we

note that accuracy values corresponding to the sports

and vegetation concepts are equal to 1. If compared

to accuracy values corresponding to the works

proposed by (Hamroun et al), we can see a significant

improvement. We observe that all the accuracy

values exceed 0.85, which means that the

improvement encompasses all the concepts. It is

broadly clear that the suggested Interactive Search

technique improves the system performance.

100

76,5

52,64

72,7

73,98

57,85

63,91

74,36

71,78

65,47

55,3

84,7

75,7

74,8

78,3

80,1

79,7

76,6

0,2

0,4

0,6

0,8

1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930

(Elleuchetal) (Wasedaetal) (Xiaoetal) LAMIRA(Hamrounetal) SVI_SIRI

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

200

Figure 14: Precision value.

5 CONCLUSION

In this paper, we propose a semantic indexing model

implemented through a new framework called SIRI,

which represents an extension of VISEN based on a

new indexing method that combines low-level and

high-level descriptors. On the basis of this

combination, SIRI becomes capable of responding to

any type of user. Specialists such as medical doctors

who search for particular areas in the videos, can

retrieve relevant results with image queries. The

system response is based on low-level descriptors.

However, if we are in the case of a non-specialist user

looking for general information on medicine, the

research can be based on other forms of queries and

some of the semantic descriptors. In future works we

will try to improve the request response time by using

new algorithms in machine learning

REFERENCES

Kundu, Malay, K., Manish, C., and Samuel, R. (2015). A

Dubey, Shiv Ram, Satish Kumar Singh, and Rajat

Kumar Singh: ’Local neighbourhood-based robust

color occurrence descriptor for colour image retrieval’,

IET Image Processing, 2015,9, (7), pp. 578–586

Wang, Xiang-Yang, Yong-Jian Yu, and Hong-Ying Yang:

’An effective image retrieval scheme using color,

texture and shape features’, Computer Standards &

Interfaces, 2011,33, (1), pp.59–68

Pass, Greg, and Ramin Zabih: ’Histogram refinement for

content-based image retrieval’, Pro-ceedings 3rd IEEE

Workshop on Applications of Computer Vision

(WACV’96), Sarasota, FL, 1996, pp. 96–102

Singha, M., Hemachandran, K. and Paul, A.: ’Content-

based image retrieval using the com-bination of the fast

wavelet transformation and the colour histogram’, IET

Image Processing, 2012,6, (9), pp. 1221–1226

Chun, Young Deok, Nam Chul Kim, and Ick Hoon Jang:

’Content-based image retrieval using multiresolution

color and texture features’, IEEE Transactions on

Multimedia, 2008,10, (6), pp. 1073–1084

Hamroun. M, Lajmi. S, Nicolas. H, Amous,VISEN: A

Video Interactive Retrieval Engine Based on Semantic

Network in large video collections, International

Database Engineering & Applications Symposium.

IDEAS 2019: 25:1-25:10.

Hamroun. M, Lajmi. S, Nicolas. H, Amous, Descriptor

optimization for Semantic Concept Detection Using

Visual Content, International Journal of Strategic

Information Technology and Applications IJSITA

10(1): 40-59 (2019).

N., Elleuch, A., Ben Ammar and A., M., Alimi. (2015). A

generic framework for semantic video indexing based

on visual concepts/contexts detection. In Mutimedia

Tools and application.

S. Fadaei, R. Amirfattahi, et M. R. Ahmadzadeh, «New

content-based image retrieval system based on

optimised integration of DCD, wavelet and curvelet

features», IET Image Processing, vol. 11, no 2, p.

89-98, 2017, doi: 10.1049/iet-ipr.2016.0542.

K. Schoemann, et al. 2010. Video browsing interfaces and

applications: a review. SPIE Reviews.

M. Ben Halima, M. Hamroun, S. Moussa and A.M. Alimi,

2013 “An interactive engine for multilingual video

browsing using semantic content”, International

Graphonomics Society Conference IGS, Nara Japan, pp

183-186.

R. Villa, N. Gildea, and J. M. Jose. 2008. Facetbrowser: a

user interface for complex search tasks. In MM ’08:

Proceeding of the 16th ACM international conference

on Multimedia, New York, NY, USA, pp 489–498.

M. Del Fabro, K. Schoeffmann, and L. Boeszoermenyi,

2010, Instant Video Browsing: A tool for fast

Nonsequential Hierarchical Video Browsing. In

Workshop of Intercative Multimedia Applications.

C.-H. Lin, R.-T. Chen, et Y.-K. Chan, « A smart content-

based image retrieval system based on color and texture

feature », Image and Vision Computing, vol. 27, no 6,

p. 658-665, mai 2009, doi: 10.1016/j.imavis.20

08.07.004.

100

120

VISEN(Hamrounetal) ISE(Hamrounetal) (BenHallimaetal) SIRI

Multimedia Indexing and Retrieval: Optimized Combination of Low-level and High-level Features

201

M. E. ElAlami, « A novel image retrieval model based on

the most relevant features », Knowledge-Based

Systems, vol. 24, no 1, p. 23-32, févr. 2011, doi:

10.1016/j.knosys.2010.06.001.

S. Murala, R. P. Maheshwari, et R. Balasubramanian,

« Local Tetra Patterns: A New Feature Descriptor for

Content-Based Image Retrieval », IEEE Transactions

on Image Processing, vol. 21, no 5, p. 2874-2886, mai

2012, doi: 10.1109/TIP.2012.2188809.

E. Yildizer, A. M. Balci, T. N. Jarada, et R. Alhajj,

« Integrating wavelets with clustering and indexing for

effective content-based image retrieval », Knowledge-

Based Systems, vol. 31, p. 55-66, juill. 2012, doi:

10.1016/j.knosys.2012.01.013.

M. K. Kundu, M. Chowdhury, et S. Rota Bulò, « A graph-

based relevance feedback mechanism in content-based

image retrieval », Knowledge-Based Systems, vol. 73,

p. 254-264, janv. 2015, doi:

10.1016/j.knosys.2014.10.009.

S. R. Dubey, S. K. Singh, et R. K. Singh, «Boosting local

binary pattern with bag-of-filters for content based

image retrieval», in 2015 IEEE UP Section Conference

on Electrical Computer and Electronics (UPCON),

2015, p. 1-6, doi: 10.1109/UPCON.2015.7456703.

S. Lefèvre, C. L'Orphelin and N. Vincent : "Extraction

multicritère de texte incrusté dans les séquences vidéo".

Colloque International sur l'Ecrit et le Document

(CIFED), 2004.

M. Worring, C. Snoek, O. de Rooji, G. P. Nguyen, R. Van

Balen and D. Koelna: "Médiamill : Advanced browsing

in news vidéo archives". CIVR, pages 533-536, 2006.

V. Stefanos, M. Anastasia, K. Paul, D. Anastasios, M.

Vasileios and K. Ioannis: "VERGE: A Video

Interactive Retriveal Engine", 2010.

P. FAUDEMAY ans C. SEYRAT: " Intelligent delivery of

personalised video programmes from a video database"

International workshop on Database anx EXpert

systems Applications, 172-177, 1997

L-I. Kuncheva, J-C. Bezdek, and R-P. Duin. Decision

templates for multiple classier fusion: An

experiemental comparaison.Pattern Recognition, 34

:299{ 314, 2001

A-K. Jain, R-P. Duin, and J. Mao. Combination of weak

classiers. IEEE Trans. on Pattern Analysis and Machine

Intelligence, 22(1), 2000

L-I. Kuncheva. Fuzzy versus nonfuzzy in combining

classiers designed by bossting. IEEE Transactions on

fuzzy systems, 11(6), 2003.

Waseda at TRECVID 2015: Semantic Indexing, Kazuya

Ueki and Tetsunori Kobayashi, TREVVID 2015.

Qualcomm Research and University of Amsterdam at

TRECVID 2015: Recognizing Concepts, Objects, and

Events in Video, C.G.M. Snoek, S. Cappallo, D.

Fontijne, D. Julian, D.C. Koelma, P. Mettes, K.E.A. van

de Sande, A. Sarah, H. Stokman, R.B. Towal,

TRECVID 2015.

S. R. Dubey, S. K. Singh, et R. K. Singh, « Rotation and

scale invariant hybrid image descriptor and retrieval »,

Computers & Electrical Engineering, vol. 46, p.

288-302, août 2015, doi: 10.1016/j.compeleceng.20

15.04.011.

Hamroun. M, Lajmi. S, Nicolas. H, Amous. I, ISE:

Interactive Image Search Using Visual Content: In

Proceedings of the 20th International Conference on

Enterprise Information Systems (ICEIS 2018) -

Volume 1, pages 253-261. ISBN: 978-989-758-298-1

(ICEIS 2018).

M. Hamroun, S. Lajmi, H. Nicolas and I. Amous, “An

Interactive Video Browsing With VINAS System”, In

Proceedings of the 15th ACS/IEEE International

Conference on Computer Systems and Applications

AICCSA 2018, Aqaba, Jordan.

N. Jhanwar, S. Chaudhuri, G. Seetharaman, et B.

Zavidovique, « Content based image retrieval using

motif cooccurrence matrix », Image and Vision

Computing, vol. 22, no 14, p. 1211-1220, déc. 2004,

doi: 10.1016/j.imavis.2004.03.026.

Y. Xiao, J. Wu, et J. Yuan, « mCENTRIST: A Multi-

Channel Feature Generation Mechanism for Scene

Categorization », IEEE Transactions on Image

Processing, vol. 23, no 2, p. 823-836, févr. 2014, doi:

10.1109/TIP.2013.2295756.

A. Irtaza et al., « An Ensemble Based Evolutionary

Approach to the Class Imbalance Problem with

Applications in CBIR », Applied Sciences, vol. 8, no 4,

p. 495, avr. 2018, doi: 10.3390/app8040495.

M. E. ElAlami, «A new matching strategy for content based

image retrieval system», Applied Soft Computing, vol.

14, p. 407-418, janv. 2014, doi: 10.1016/j.asoc.20

13.10.003.

R. Ashraf, K. Bashir, A. Irtaza, et M. T. Mahmood,

«Content Based Image Retrieval Using Embedded

Neural Networks with Bandletized Regions», Entropy,

vol. 17, no 6, p. 3552-3580, juin 2015, doi:

10.3390/e17063552.

E. Walia et A. Pal, «Fusion framework for effective color

image retrieval», Journal of Visual Communication and

Image Representation, vol. 25, no 6, p. 1335-1348, août

2014, doi: 10.1016/j.jvcir.2014.05.005.

S. Zeng, R. Huang, H. Wang, et Z. Kang, «Image retrieval

using spatiograms of colors quantized by Gaussian

Mixture Models», Neurocomputing, vol. 171, p.

673-684, janv. 2016, doi: 10.1016/j.neucom.20

15.07.008.

Y. Zhou, F.-Z. Zeng, H. Zhao, P. Murray, et J. Ren,

«Hierarchical Visual Perception and Two-Dimensional

Compressive Sensing for Effective Content-Based

Color Image Retrieval», Cogn Comput, vol. 8, no 5, p.

877-889, oct. 2016, doi: 10.1007/s12559-016-9424-6.

N. Shrivastava et V. Tyagi, «An efficient technique for

retrieval of color images in large databases»,

Computers & Electrical Engineering, vol. 46, p.

314-327, août 2015, doi: 10.1016/j.compeleceng.20

14.11.009

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

202