Department of Computer Science

Permanent URI for this communityhttps://dspace.univ-soukahras.dz/handle/123456789/15

Browse

Now showing 1 - 20 of 71

3D Mobility, Resizing and Mobile Sink Nodes in Reconfigurable Wireless Sensor Networks based on Multi-agent Architecture under Energy Harvesting Constraints
(SCITEPRESS, 2020-01) Hanene Rouainia; Hanen Grichi; Laid Kahloul; Mohamed Khalgui
This paper deals with reconfigurable wireless sensor networks (RWSNs) to be composed of a set of sensor nodes, which monitor the physical and chemical conditions of the environment. RWSNs adapt dynamically their behaviors to their environment. The main challenge in RWSN is to keep the network alive as long as possible. We apply a set of solutions for energy problems by using 3D mobility, resizing and mobile sink nodes. These solutions are based on a multi-agent architecture employing a wireless communication protocol. Moreover, we develop an application named RWSNSim that allows us to simulate an RWSN and apply the proposed solutions. The performance of the proposed approach is demonstrated through a case study. The case study consists of surveying of fire in a forest which is simulated with RWSNSim application.
A modified incremental density based clustering algorithm
(IEEE, 2022) Aida Chefrour
Cluster analysis, generally known as clustering, is a technique for separating data into groups (clusters) of similar objects. Except if the system is completely retrained, traditional clustering classifiers will be unable to learn new information and knowledge (attributes, examples, or classes). Only incremental learning, which outperforms when new data objects are introduced into an existing database, can solve this problem. These evolutionary strategies are applied to dynamic databases by updating the data. We’ll choose to study the Incremental Density- Based Spatial Clustering of Applications with Noise algorithm because of its capacity to discover arbitrary clusters and identify noise. In this study, a modified version of the Incremental Density Based Clustering Algorithm using an Adaptive Median Filtering Technique was used. The difference between our previous proposed AMF-IDBSCAN and the proposed algorithm developed in this work is in the evaluation performance stage. The key idea consists of a database change in the case of introducing new data items to an existing database in order to improve performance. We conducted several experiments on benchmark and synthetic data collected from the University of California Irvine repository in terms of the Generalized Dunn Index, Davies Bouldin Index, and change of time (milliseconds) with the increment of data in the original database. Experiments with datasets of various sizes and dimensions show that the proposed algorithm enhances clustering when compared to several current incremental wellknown techniques.
A New Simulation Tool for Sensor Networks Based on an Energy-Efficient and Fault-Tolerant Methodology
(Springer, Cham, 2023-07) Hanene Rouainia; Hanen Grichi; Laid Kahloul; Mohamed Khalgui
Recently, reconfigurable wireless sensor networks (RWSNs) have attracted a lot of attention in research and industrial communities. They became more complex and dynamic systems which led to the emergence of many challenges. The lack of energy, real-time constraints, and software and hardware failures are the most important challenges in RWSNs. Indeed, several solutions have proposed to come up with these challenges. To avoid huge costs in terms of money, time, and effort of real experimentation, networks’ simulation tools have become an essential necessity to study the impact of the proposed solutions. In this work, we propose a new energy-efficient and fault-tolerant methodology that composed of a set of solutions summarized in the use of mobile sink nodes (MSNs), application of the mobility, resizing, and test packet technique using a multi-agent architecture and an energy-efficient routing protocol. Moreover, we propose a new discrete-event simulation tool named RWSNSim designed for sensor networks (WSNs & RWSNs). We present its description, modeling, and provided services. The proposed simulation tool allows simulating sensor networks with and without application of the proposed methodology. Finally, we simulate a case study using RWSNSim in a 3D environment which proves the effectiveness of the proposed methodology and demonstrate the efficiency of the suggested simulation tool.
A Novel Incremental Learning Algorithm Based on Incremental Vector Support Machina and Incremental Neural Network Learn++
(Lavoisier, 2019) Aida Chefrour; Labiba Souici-Meslati; Iness Difi; Nesrine Bakkouche
Incremental learning refers to the learning of new information iteratively without having to fully retain the classifier. However, a single classifier cannot realize incremental learning if the classification problem is too complex and scalable. To solve the problem, this paper combines the incremental support vector machine (ISVM) and the incremental neural network Learn++ into a novel incremental learning algorithm called the ISVM-Learn++. The two incremental classifiers were merged by parallel combination and weighted sum combination. The proposed algorithm was tested on three datasets, namely, three databases Ionosphere, Haberman's Survival, and Blood Transfusion Service Center. The results show that the ISVM Learn ++ achieved a learning rate of 98 %, better than that of traditional incremental learning algorithms. The research findings shed new light on incremental supervised machine learning.
A transfer learning-basedrecommender system
(2024) AYA SAHRAOUI
Transformers have revolutionized the field of Automatic Data Processing Natural Language (NLP) producing remarkable results In this study, we propose an extension of this success by introducing A new recommendation system based on Transformers. We exploited their ability to capture complex relationships between data to provide insights precise and personalized recommendations. Using extracted data from Amazon, we conducted an in-depth experiment to evaluate the effectiveness of our recommendation model. The results demonstrated the relevance and effectiveness of our approach, highlighting the advantage of transfer learning in improving recommendation performance. This research opens new perspectives in the field of systems of recommendation by exploiting recent advances in NLP to provide more precise and relevant recommendations to users. ------------------------------------------------------------------------------ لقد أحدثت المحولات ثورة في مجال المعالجة التلقائية للبيانات اللغة الطبيعية (NLP) تحقق نتائج ملحوظة في هذه الدراسة، نقترح تمديد هذا النجاح من خلال تقديم جديد نظام التوصية على أساس المحولات. لقد استغلناهم القدرة على التقاط العلاقات المعقدة بين البيانات لتقديم رؤى توصيات دقيقة وشخصية. باستخدام البيانات المستخرجة من أمازون، أجرينا تجربة متعمقة للتقييم فعالية نموذج التوصية لدينا. وأظهرت النتائج أهمية وفعالية نهجنا، وتسليط الضوء على ميزة نقل التعلم في تحسين أداء التوصية. يفتح هذا البحث آفاقا جديدة في مجال نظم التوصية من خلال استغلال التطورات الحديثة في البرمجة اللغوية العصبية لتقديمها توصيات أكثر دقة وذات صلة للمستخدمين. ------------------------------------------------------------------------------ Les Transformers ont révolutionné le domaine du Traitement Automatique du Langage Naturel (NLP) en produisant des résultats remarquables Dans cette étude, nous proposons une extension de cette réussite en introduisant UN nouveau système de recommandation basé sur les Transformers. Nous avons exploité leur capacité à capturer les relations complexes entre les données pour proposer des recommandations précises et personnalisées. En utilisant des données extraites d'Amazon, nous avons mené une expérimentation approfondie pour évaluer l'efficacité de notre modèle de recommandation. Les résultats ont démontré la pertinence et l'efficacité de notre approche, mettant en évidence l'avantage du transfert d'apprentissage dans l'amélioration des performances de recommandation. Cette recherche ouvre de nouvelles perspectives dans le domaine des systèmes de recommandation en exploitant les avancées récentes en NLP pour fournir des recommandations plus précises et pertinentes aux utilisateurs.
Adaptation with Four Dimensional Personalization Criteria Based on Felder Silverman Model
(IGI GLOBAL, 2017) Samia Drissi; Abdelkrim Amirat
In the past decades, various systems have been proposed to provide students with a better learning environment by taking personal factors into account. Learning styles have been one of the widely adopted factors in the previous studies as a reference for adapting learning content or organizing the content. However, very few researchers give an idea of matching e-media with appropriate teaching and learning styles and very few studies give an idea of which appropriate combinations of electronic media and learning styles are more effective than other. In this paper, the authors aim to prototype an AFDPC-FS system (Adaptation with Four Dimensional Personalization Criteria based on Felder Silverman model). Their system presents a general framework for combining and adapting teaching strategies, learning styles and electronic media according to Felder-Silverman’slearning style model. An experiment was designed to explore the effect of adaptation to different learning styles when learning materials were matched with learning styles. In particular it was set up to see whether there are significant differences in learning achievement and cognitive load between two groups, an experimental group who studied with learning style-fit version and a control group who studied with non-fit version of the system without adaptation to learning styles. The experimental results showed that the proposed system could improve the learning achievements of the students. Moreover, it was found that the students’ cognitive load was significantly decreased
AMF-IDBSCAN: Incremental Density Based Clustering Algorithm using Adaptive Median Filtering Technique
(Slovene Society Informatika, 2019) Aida Chefrour; Labiba Souici-Meslati
Density-based spatial clustering of applications with noise (DBSCAN) is a fundamental algorithm for density-based clustering. It can discover clusters of arbitrary shapes and sizes from a large amount of data, which contains noise and outliers. However, it fails to treat large datasets, outperform when new objects are inserted into the existing database, remove noise points or outliers totally and handle the local density variation that exists within the cluster. So, a good clustering method should allow a significant density modification within the cluster and should learn dynamics and large databases. In this paper, an enhancement of the DBSCAN algorithm is proposed based on incremental clustering called AMF-IDBSCAN which builds incrementally the clusters of different shapes and sizes in large datasets and eliminates the presence of noise and outliers. The proposed AMF-IDBSCAN algorithm uses a canopy clustering algorithm for pre-clustering the data sets to decrease the volume of data, applies an incremental DBSCAN for clustering the data points and Adaptive Median Filtering (AMF) technique for post-clustering to reduce the number of outliers by replacing noises by chosen medians. Experiments with AMF-IDBSCAN are performed on the University of California Irvine (UCI) repository UCI data sets. The results show that our algorithm performs better than DBSCAN, IDBSCAN, and DMDBSCAN.
An Adaptive E-Learning System based on Student’s Learning Styles: An Empirical Stud
(IGI GLOBAL, 2016) Samia Drissi; Abdelkrim Amirat
Personalized e-learning implementation is recognized as one of the most interesting research areas in the distance web-based education. Since the learning style of each learner is different one must fit e-learning with the different needs of learners. This paper presents an approach to integrate learning styles into adaptive e-learning hypermedia. The main objective was to develop a new Adaptive Educational Hypermedia System based on Honey and Mumford learning style model (AEHS-H&M) and assess the effect of adapting educational materials individualized to the student’s learning style. To achieve the main objectives, a case study was developed. An experiment between two groups of students was conducted to evaluate the impact on learning achievement. Inferential statistics were applied to make inferences from the sample data to more general conditions was designed to evaluate the new approach of matching learning materials with learning styles and their influence on student’s learning achievement. The findings support the use of learning styles as guideline for adaptation into the adaptive e-learning hypermedia systems;
An Intersection Attack on the CirclePIN Smartwatch Authentication Mechanism
(IEEE Internet of Things Journal, 2024-04-01) Djalel Chefrour; Yasser Sedira; Samir Chabbi
We present a thorough security analysis of a recent smartwatch authentication mechanism called CirclePIN, which was considered resilient to several attacks, including shoulder surfing and video recording. This mechanism avoids the direct entry of the personal identification number (PIN) by using consecutive screens of random colors that fool the attacker. We disclose a vulnerability in CirclePIN inherent to the way in which the users match the random colors to their PINs’ digits and we illustrate how to exploit it with an intersection attack. This attack uses the information extracted from multiple video recordings of legitimate authentication sessions. We prove that it has a high probability of revealing the user PIN with only three video recordings and always succeeds with five. Our proof is twofold. We formulate the theoretical probability of success for the attack as a function of the number of available video recordings. Then, we validate this formula with a simulation of a large number of attacks to compute their experimental probability of success. In our estimation, manual information extraction takes around 1 min per exploitable video recording. So, a complete intersection attack is cost effective in terms of time, as it lasts 5 min or less.
An Intrusion Detection System Based on Deep Learning and Genetic Algorithm
(2023) Sabrina Djeddou
Cyber security is of very importance in today's interconnected world. It ensures the protection of data, preserves privacy, and more. Intrusion Detection Systems play a crucial role in cyber security by actively monitoring network traffic and system activities. On the other hand we have deep learning, which is a subfield of machine learning that uses artificial neural networks. Deep learning could be used to detect unusual patterns or behavior within a network that may indicate a security threat. When it comes to optimizing deep learning models, hyperparameter tuning is an important step that can have a significant impact on their performance. Exploring all hyperparameter combinations can be impractical or impossible due to their vast number of possibilities. In hyperparameter tuning, metaheuristics like Genetic Algorithms can leads to better solutions that would be difficult or time-consuming to obtain through manual optimization. In this work we utilized a convolutional neural network 2D, leveraging its outstanding performance across various domains. By optimizing it with genetic algorithms, we surpassed other tested algorithms with remarkable results. ---------------------------------------------------------------------------------------------- للأمن المعلوماتي أهمية كبيرة في عالم اليوم المترابط. يضمن حماية البيانات ؛ ويحافظ على الخصوصية ؛ وأكثر من ذلك. تلعب أنظمة كشف التسلل دورًا مهمًا في الأمن المعلوماتي من خلال المراقبة النشطة لحركة مرور الشبكة وأنشطة النظام. من ناحية أخرى ؛ لدينا التعلم العميق ؛ وهو حقل فرعي من التعلم الآلي الذي يستخدم الشبكات العصبية الاصطناعية. يمكن استخدام التعلم العميق لاكتشاف الأنماط أو السلوك غير المعتاد داخل الشبكة والذي قد يشير إلى وجود تهديد أمني. عندما يتعلق الأمر بتحسين نماذج التعلم العميق ؛ يعدالضبط الفائق خطوة مهمة يمكن أن يكون لها تأثير كبير على أدائها. يمكن أن يكون استكشاف جميع مجموعات المعلمات الفائقة أمرًا غير عملي أو مستحيلا بسبب العدد الهائل من الاحتمالات. في ضبط المعلمات الفائقة ؛ يمكن أن تؤدي الخصائص الوصفية مثل الخوارزميات الجينية إلى حلول أفضل قد تكون صعبة أوتستغرق وقتا طويلا من خلال التحسين اليدوي. في هذا العمل ‎٠‏ استخدمنا شبكة عصبية تلافيفية ثنانية الأبعاد ؛ مستفيدين من أدائها المتميز في مختلف المجالات. من خلال تحسينها باستخدام الخوارزميات الجينية ؛ تجاوزنا الخوارزميات المختيرة الأخرى بنتائج ملحوظة. ---------------------------------------------------------------------------------------------- La sécurité informatique revêt une très grande importance dans le monde interconnecté d'aujourd'hui. Elle garantit la protection des données, préserve la vie privée, et bien plus encore. Les systèmes de détection d'intrusion jouent un rôle crucial dans la sécurité informatique en surveillant activement le trafic réseau et les activités du système. D'un autre côté, nous avons l'apprentissage profond, qui est un sous-domaine de l'apprentissage automatique utilisant des réseaux neuronaux artificiels. L'apprentissage profond peut être utilisé pour détecter des schémas ou des comportements inhabituels au sein d'un réseau pouvant indiquer une menace pour la sécurité. En ce qui concerne l'optimisation des modèles d'apprentissage profond, l'accord des hyperparamètres est une étape importante pouvant avoir un impact significatif sur leurs performances. Explorer toutes les combinaisons d'hyperparamètres peut être impraticable ou impossible en raison de leur nombre considérable de possibilités. Dans l'accord des hyperparamètres, des métaheuristiques comme les algorithmes génétiques peuvent conduire à de meilleures solutions qui seraient difficiles ou longues à obtenir par une optimisation manuelle. Dans ce travail, nous avons utilisé un réseau neuronal convolutif 2D, en tirant parti de ses performances exceptionnelles dans divers domaines. En l'optimisant avec des algorithmes génétiques, nous avons dépassé les autres algorithmes testés avec des résultats remarquables.
An Intrusion Detection System for the Internet of Things based on a hybrid federated learning model
(2024) Zerafa Marwa
Abstract In recent years, the rapid expansion of the Internet of Things (IoT) has necessitated robust and adaptive security measures to protect against increasingly sophisticated cyber threats. This dissertation presents a comprehensive analysis of federated learning (FL) approaches for intrusion detection in IoT environments. The focus is on developing a decentralized, privacy- preserving intrusion detection system (IDS) that leverages federated learning frameworks to enhance security without compromising data privacy. Various deep learning models, including Convolutional Neural Networks (CNNs),Deep Neural Networks (DNNs) and Long Short-Term Memory (LSTM) networks, are evaluated for their effectiveness in the proposed Hybrid Feder- ated Intrusion Detection System (HybFed IDS).Extensive experiments using real-world IOT datasets such as CICIOT2023 demonstrate the efficacy of these models, achieving high accu- racy and detection rates. The findings highlight the potential of FL-based IDS to offer scalable, resilient, and secure solutions for protecting IoT networks from cyber threats. ---------------------------------------------------------------------------------------- في السنوات الأخيرة، استدعى التوسع السريع لإنترنت الأشياء اتخاذ تدابير أمنية قوية ومتكيّفة لحمايتها من التهديدات السيبرانية المتزايدة التعقيد. تقدم هذه الرسالة تحليلاً شاملاً لأساليب التعلم الموحد للكشف عن التطفل في بيئات إنترنت الأشياء. يركز البحث على تطوير نظام كشف التطفل لامركزي يحافظ على الخصوصية، ويعتمد على أطر التعلم الموحد لتعزيز الأمان دون المساس بخصوصية البيانات. يتم تقييم نماذج التعلم العميق المختلفة، بما في ذلك الشبكات العصبية التلافيفية والشبكات العصبية العميقة وشبكات الذاكرة الطويلة القصيرة المدى، لفعالية نظام الكشف عن التسلل الموحد الهجين المقترح. وتظهر التجارب الموسعة باستخدام مجموعات بيانات إنترنت الأشياء الحقيقية فعالية هذه النماذج، محققة دقة عالية ومعدلات كشف عالية.تبرز النتائج إمكانيات نظام الكشف عن التطفل القائم على التعلم الموحد في تقديم حلول قابلة للتطوير، ومرنة، وآمنة لحماية شبكات إنترنت الأشياء من التهديدات السيبرانية ----------------------------------------------------------------------------------------- Ces dernières années, l’expansion rapide de l’Internet des objets (IoT) a nécessité des solutions robustes. et des mesures de sécurité adaptatives pour se protéger contre les cybermenaces de plus en plus sophistiquées. Cette thèse présente une analyse complète des approches d’apprentissage fédéré (FL) pour détection d’intrusion dans les environnements IoT. L’accent est mis sur le développement d’un système décentralisé de protection de la vie privée. préserver le système de détection d’intrusion (IDS) qui exploite les cadres d’apprentissage fédéré pour améliorez la sécurité sans compromettre la confidentialité des données. Divers modèles d’apprentissage profond, notamment Réseaux de neurones convolutifs (CNN), réseaux de neurones profonds (DNN) et long terme Les réseaux de mémoire (LSTM) sont évalués pour leur efficacité dans la fédération hybride proposée. Système de détection d’intrusion (HybFed IDS). Expériences approfondies utilisant l’IOT du monde réel des ensembles de données tels que CICIOT2023 démontrent l’efficacité de ces modèles, permettant d’obtenir une précision élevée. taux de racée et de détection. Les résultats mettent en évidence le potentiel de l’IDS basé sur FL à offrir des solutions évolutives et des solutions résilientes et sécurisées pour protéger les réseaux IoT contre les cybermenaces.
An Optimized Path Planning for Wheeled Robot in Obstacle Environments
(IEEE explore, 2021-11-03) Mohamed Amine Yakoubi; Djamel Bektache; Abderahmane Gaham; Raouf Toumi
In this paper, we have proposed an obstacle avoidance algorithm for a path planning in unknown environment for a mobile robot based on the fuzzy logic control. Therefore, the wheeled mobile robot is equipped with 3 wheels, one steering wheel and two fixed wheels and mounted on the same axis. Its task is to move from a starting position to a target position. For this, our proposed algorithm creates one or more imaginary target and applies a fuzzy logic control system, which is adopted by a rule table that is induced from two inputs data (the distance and the angle between the robot and the target) and two outputs data (the angle orientation and velocity of the steer wheel). Experimental results show the effectiveness of the proposed algorithm.
Application et modélisation d'un protocole de communication pour la sécurité routière
(University Badi Mokhtar Annaba, 2014-10-25) BEKTACHE DJAMEL Chérif Tolba , Ghoualmi Nacera
Application mobile pour la recherche des métiers Dans le secteur du BTP
(2024) Aya Hefaidia
Our mobile app for job search in BTP has been designed to make it easier to find jobs in the construction and public works industry. It aims to connect construction professionals with clients who offer job opportunities that match candidates' skills and qualifications. Through this application, customers will be able to find capable workers in the construction field quickly and efficiently, and workers can apply to offers that suit their skills and professional interests. The main objective is to provide users with a user-friendly and efficient platform to simplify the process of finding a job in the construction industry. The methodology adopted is based on the in-depth analysis of user needs, the design of an intuitive interface and the implementation of key functionalities such as searching for jobs and workers, consulting worker profiles, and applying for jobs. The use of modern technologies such as Flutter for cross-platform development and Firebase for data storage ensures a smooth and secure experience for users. In summary, the app offers an innovative solution to make it easier to find jobs in the construction industry, while providing customers with an efficient way to find qualified candidates. --------------------------------------------------------------------------------------- تم تصميم تطبيق الهاتف المحمول الخاص بنا لتسهيل البحث عن عمل في مجال البناء والأشغال العامة. حيث يهدف إلى ربط العمال المتخصصين في هذا المجال بالعملاء الذين يقدمون فرص عمل تتناسب مع مهارات المترشحين ومؤهلاتهم. بفضل هذا التطبيق يمكن للعملاء إيجاد عمال متمكنون في مجال البناء والأشغال العامة بطريقة سريعة، ويمكن للعمال التقدم بطلب للحصول على وظيفة التي تتماشى مع مهاراتهم واهتماماتهم المهنية. حيث يكمن الهدف الرئيسي في تزويد المستخدمين بمنصة سهلة الاستخدام وفعالة لتبسيط عملية البحث عن عمل في مجال البناء و التشييد. تعتمد المنهجية المعتمدة على التحليل العميق لاحتياجات المستخدم، وتصميم واجهة بديهية لتنفيذ الوظائف الرئيسية والبحث عن الوظائف والأعمال، وتصفح ملفات تعريف العمال و التقديم للوظائف. يضمن استخدام التقنيات الحديثة مثل فلاتر للتطوير عبر الأنظمة الأساسية و فاير بايزلتخزين البيانات تجربة سلسة وآمنة للمستخدمين. باختصار، فإن التطبيق يقدم حلا مبتكرا لتسهيل العثور على وظائف في مجال البناء ، مع تزويد العملاء بطريقة فعالة للعثور على المرشحين المؤهلين. --------------------------------------------------------------------------------------- Notre application mobile pour la recherche d'emploi en BTP a été conçue pour faciliter la recherche d'emplois dans l'industrie de la construction et des travaux publics. Elle vise à connecter les professionnels du BTP avec les clients qui offrent des opportunités d'emploi qui correspondent aux compétences et aux qualifications des candidats. Grâce à cette application, les clients pourront trouver des travailleurs capables dans le domaine du BTP de manière rapide et efficace, et les travailleurs peuvent appliquer à des offres qui suivent leurs compétences et leurs intérêts professionnels. L'objectif principal est de fournir aux utilisateurs une plateforme conviviale et efficace pour simplifier le processus de recherche d'emploi dans le secteur du BTP. La méthodologie adoptée se base sur l’analyse approfondie des besoins des utilisateurs, la conception d'une interface intuitive et la mise en œuvre de fonctionnalités clés telles que la recherche d'emplois et des travailleurs, la consultation de profils des travailleurs, et la postulation des travaux. L'utilisation de technologies modernes telles que Flutter pour le développement cross-Platform et Firebase pour le stockage des données garantit une expérience fluide et sécurisée pour les utilisateurs. En résumé, l'application offre une solution innovante pour faciliter la recherche d'emplois dans le secteur du BTP, tout en offrant aux clients un moyen efficace de trouver des candidats qualifiés.
Audit and securing the website of the University of Souk Ahras
(2023) Abdallah Amirat
Information, or the golden bit, has become the greatest source of power in the world today. And it's making people envious - everyone wants to get their hands on it, especially the Hackers (the Black Hats), network and systems experts who use their talents to break into the heart of our organizations and steal our data, destroying our information by injecting malware and vandalizing code. This master is a contribution to the enhancement of the University web site security through: the creation of a mirror site for the purpose of testing and development, the upgrade of the software packages used by the web site and the audit of the latter with state-of-the-art tools. Namely, we used the Zed Attack Proxy audit tool to detect and analyze the security vulnerabilities of the website that allow attacks such as Cross-Site Scripting (XSS). We then worked on immediate changes to the site to patch certain vulnerabilities and prevent possible attacks.
Background modeling using deep learning
(2024) Wassim Boulouh; Mohamed El Bachir Boubaidja
Background subtraction plays a pivotal role in computer vision applications, particularly in video surveillance, where accurate detection of moving objects in variable environmental conditions is paramount. This report presents a robust background subtraction system using deep learning. The system begins with a pre-processing stage, where the video frames are standardized and the noise is reduced through bilateral filtering. Ground truth images also undergo the similar pre-processing steps in order to align them with the original data. In the processing stage, we have used a modified U-net architecture as tools for pixel segmentation. The modification introduced on the U-net architecture, including an additional convolution layer in the encoder part, which enhance feature extraction and improve model performance, particularly for larger and more complex images. After the processing step, binary images are generated. The post-processing steps involve morphological operations such as dilation and erosion to refine the binary images, correcting false detections and enhancing accuracy. The test of our model on public dataset, demonstrates the performance of our proposition. --------------------------------------------------------------------------------- La soustraction d’arrière-plan joue un rôle crucial dans les applications de vision par ordinateur, en particulier dans la vidéosurveillance, où la détection précise des objets en mouvement dans des conditions environnementales variables est primordiale. Ce rapport présente une méthode de soustraction de fond robuste pour les systèmes de vidéosurveillance. Le système commence par une étape de prétraitement, où les images du vidéo sont standardisées et le bruit est réduit par le filtre bilatéral. Les images de vérité de terrain subissent également des étapes de prétraitement similaires pour les aligner avec les données d’origine. Notre modèle proposé utilisait une architecture U-net modifiée comme outils pour la segmentation des pixels. La modification introduite sur l’architecture U-net, y compris une couche convolutive supplémentaire dans la partie encodeur, qui améliore l’extraction des caractéristiques et les performances du modèle, en particulier pour les images plus grandes et plus complexes. Après l’étape de traitement, des images binaires sont générées, distinguant les pixels d’arrière-plan et d’avant-plan. Les étapes de post-traitement impliquent des opérations morphologiques telles que la dilatation et l’érosion pour affiner les images binaires, corriger les fausses détections et améliorer la précision. Le test de notre modèle sur jeu de données publiques, démontre la performance de notre proposition. --------------------------------------------------------------------------- تلعب عملية طرح الخلفية دورًا حاسمًا في تطبيقات رؤية الحاسوب، خاصة في مجال المراقبة بالفيديو، حيث يكون التعرف الدقيق على الأجسام المتحركة في ظروف بيئية متغيرة أمرًا أساسيًا. يقدم هذا التقرير طريقة قوية لطرح الخلفية لأنظمة المراقبة بالفيديو. يبدأ النظام بمرحلة معالجة مسبقة، حيث تُقيس الصور الفيديوية وتُقلل الضوضاء باستخدام فلتر ثنائي المعالم. تخضع صور الحقيقة الميدانية أيضًا لخطوات معالجة مسبقة مماثلة لمزامنتها مع البيانات الأصلية. استخدمت نموذجنا المقترح بنية U-netمعدلة كأداة لتقسيم البكسلات. تتضمن التعديلات المُجراة على بنية U-netإضافة طبقة تحويلية إضافية في جزء المُشفر، مما يعزز استخراج السمات وأداء النموذج، خاصة للصور الأكبر حجمًا والأكثر تعقيدًا. بعد مرحلة المعالجة، يتم إنشاء صور ثنائية تمييز بكسلات الخلفية والأمامية. تشمل خطوات المعالجة اللاحقة عمليات مورفولوجية مثل التوسيع والتآكل لتنقية الصور الثنائية، وتصحيح الكشفات الزائفة وتحسين الدقة. اختبار نموذجنا على مجموعة بيانات عامة يبرهن على أداء اقتراحنا
CAE-CNN: Image Classification Using Convolutional AutoEncoder Pre-Training
(2022) Aida Chefrour; Samia Drissi
The work presented in this paper is in the general framework of classification using deep learning and, more precisely, that of convolutional neural networks (CNN). In particular, the convolutional autoencoder proposes an alternative for the processing of high-dimensional data, to facilitate their classification. In this paper, we propose the incorporation of convolutional autoencoders as a general unsupervised learning data dimension reduction method for creating robust and compressed feature representations in order to improve CNN performance on image classification tasks. For prediction reasons, we applied the two methods to the MNIST image databases. The use of CNN with the convolutional autoencoder gives better results compared to the individual use of each of them in terms of accuracy, to obtain a good classification of the data high-dimensional entrance.
Clusteringprofond avec auto-encodeurs convolutionnels
(2023) Silya kheraifi
Clustering is a fundamental problem in many data-driven domains. The performance of clustering heavily relies on the quality of data representation. In recent years, numerous studies have focused on using deep neural networks to learn representations that enhance clustering and lead to significant improvements in clustering performance. In this study, we utilized a convolutional auto-encoder based neural network and the classical k-means clustering method to learn better data representations that facilitate clustering. To evaluate the results, we used the MNIST database, which contains handwritten digits, and measured the performance using the normalized mutual information (NMI) and the unsupervised clustering accuracy ACC and adjusted Rand index ARI. The obtained results demonstrate that classification based on the convolutional auto-encoder outperforms clustering performed with the classical k-means algorithm. -------------------------------------------------------------------------------------- التجميع هو مشكلة أساسية في العديد من المجالات التي تعتمد على البيانات. يعتمد أداء التجميع بشكل كبير على جودة تمثيل البيانات. في السنوات الأخيرة، ركزت العديد من الدراسات على استخدام الشبكات العصبية العميقة لتعلم التمثيلات التي تعزز التجميع وتؤدي إلى تحسينات كبيرة في أداء التجميع. في هذه الدراسة، استخدمنا شبكة عصبية قائمة على التشفير التلقائي وطريقة تجميع الوسائل الكلاسيكيةk-means)(لتعلم تمثيلات بيانات أفضل تسهل التجميع. لتقييم النتائج، استخدمنا قاعدة بيانات MNIST ، التي تحتوي على أرقام مكتوبة بخط اليد ، وقمنا بقياس الأداء باستخدام مؤشر تقييم المعلومات المتبادلة (NMI)ودقة التجميع غير المشروفة ACCو مؤشر راند المعدلARI . توضح النتائج المتحصل عليها أن التصنيف المعتمد على المشفر التلقائي العميق يتفوق على التجميع المنفذ باستخدام خوارزمية الوسائل التقليدية (k-means). -------------------------------------------------------------------------------------- Le clustering est un problème essentiel dans de nombreux domaines où les données sont prédominantes. Les performances de clustering dépendent fortement de la qualité de la représentation des données. Récemment, de nombreuses études se sont concentrées sur l'utilisation de réseaux de neurones profonds pour apprendre des représentations améliorant le regroupement et entraînant une amélioration significative des performances de clustering. Dans cette étude, nous avons utilisé un réseau neuronal à base d'auto-encodeur convolutionnel ainsi que la méthode classique de regroupement k-means pour apprendre de meilleures représentations de données qui facilitent le regroupement. Pour évaluer les résultats, nous avons utilisé la base de données MNIST, qui contient des chiffres manuscrite et mesuré les performances à l'aide de l'indice d'information mutuelle normalisée (NMI) et de l'exactitude du regroupement non supervisé ACC etl’indice Rand ajustéARI. Les résultats obtenus montrent que la classification basée sur l’auto-encodeur convolutif est plus performante que le regroupement réalisé avec l'algorithme classique k-means.
Computer Vision-Based Waste Management System
(2024) Mallek Rahma RADOUANE
With the development of demographics and urbanization worldwide, waste generation rates are increasing, making its volume a worrying threat that causes the deterioration of human health and the environment. To address this problem, our research proposes a system for recognizing different types of waste based on convolutional neural network models (CNN, VGG16...). On the other hand, separating waste into several components is one of the most important steps in waste management, and this process is usually done manually by sorting. To simplify this process, we proposed a waste segmentation model using (YOLO). Both models were trained on our custom dataset to recognize 12 different types of waste. The developed system showed promising results, demonstrating high accuracy in recognizing 12 types of waste. The VGG16 model achieved an accuracy of up to 98%, while the YOLO model achieved an average accuracy of 82%. These achievements confirm the effectiveness of the system in recognizing waste in real-world conditions, contributing to improved waste recognition and separation. ----------------------------------------------------------------------------------------- مع تطور التركيبة السكانية والتحضر في جميع أنحاء العالم، تتزايد معدلات توليد النفايات، مما يجعل حجمها تهديدًا مقلقًا يتسبب في تدهور صحة الإنسان والبيئة. لمعالجة هذه المشكلة، يقترح بحثنا نظامًا للتعرف على الأنواع المختلفة للنفايات بالاعتماد على نماذج الشبكة العصبية التلافيفية (CNN, VGG16...) ومن جهة أخرى، يُعد فصل النفايات إلى عدة مكونات من أهم الخطوات في إدارة النفايات، وعادة ما تتم هذه العملية يدويًا عن طريق الفرز. لتبسيط هذه العملية، اقترحنا نموذجًا لتجزئة النفايات باستخدام (YOLO). تم تدريب النموذجين على مجموعة بيانات خاصة بنا للتعرف على 12 نوعًا مختلفًا من النفايات. أظهر النظام المطور نتائج مبشرة، حيث أظهر دقة عالية في التعرف على 12 نوعًا من النفايات، فقد حقق نموذج VGG16 دقة تصل إلى 98%، في حين حقق نموذج YOLO متوسط دقة وصل إلى 82%. تؤكد هذه الإنجازات فعالية النظام في التعرف على النفايات في أرض الواقع، مما يسهم في تحسين جودة التعرف على النفايات وفصلها. ----------------------------------------------------------------------------------------- Avec le développement de la démographie et de l’urbanisation dans le monde entier, les taux de production de déchets augmentent, rendant leur volume une menace préoccupante qui cause la détérioration de la santé humaine et de l’environnement. Pour résoudre ce problème, notre recherche propose un système de reconnaissance des différents types de déchets basé sur des modèles de réseau de neurones convolutifs (CNN, VGG16...). D’autre part, la séparation des déchets en plusieurs composants est l’une des étapes les plus importantes de la gestion des déchets, et ce processus se fait généralement manuellement par tri. Pour simplifier ce processus, nous avons proposé un modèle de segmentation des déchets utilisant (YOLO). Les deux modèles ont été entraînés sur notre propre jeu de données pour reconnaître 12 types de déchets différents. Le système développé a montré des résultats prometteurs, démontrant une haute précision dans la reconnaissance de 12 types de déchets. Le modèle VGG16 a atteint une précision allant jusqu’à 98%, tandis que le modèle YOLO a atteint une précision moyenne de 82%. Ces réalisations confirment l’efficacité du système dans la reconnaissance des déchets dans des conditions réelles, contribuant à améliorer la reconnaissance et la séparation des déchets.
Conception et développement d’un système de gestion de transferts universitaires
(2023) Rym Mellal
In the personal and academic background, the student and the university services have an essential role. Among these university services we have transfers, between universities (external) or faculties (internal), with the help of these transfers, the student will have the possibility of changing the place or the specialty of his university studies. To access this service, we will have to generate a platform where the student will follow the procedures to make a transfer request. --------------------------------------------------------------------------------------- في الخلفية الشخصية والأكاديمية ، يكون للطالب والخدمات الجامعية دور أساسي. من بين هذه الخدمات الجامعية لدينا تحويلات ، بين الجامعات (الخارجية) أو الكليات (الداخلية) ، وبمساعدة هذه التحويلات ، سيكون لدى الطالب إمكانية تغيير مكان أو تخصص دراسته الجامعية. للوصول إلى هذه الخدمة ، سيتعين علينا إنشاء منصة حيث سيتبع الطالب الإجراءات لتقديم طلب نقل --------------------------------------------------------------------------------------- Au parcours éducatif et personnel, l'étudiant et les services universitaires présentent un rôle essentiel. Parmi ces services universitaires on a les transferts, entre les universités (externes) ou les facultés (internes), à l’aide de ces transferts, l’étudiant aura la possibilité de changer le lieu ou la spécialité de ses études universitaires. Pour accéder à ce service, on va devoir générer une plateforme où l’étudiant va poursuivre des procédures pour faire une demande de transfert.