بحوث سكوبس — قصي منير دياب الدره

Semantic image retrieval aims to find images with similar content or meaning rather than just visual similarity. CNN-based feature hashing and clustering techniques have proven effective in achieving fast and accurate image retrieval by capturing semantic representations in compact forms. However, existing methods often suffer from limited semantic preservation in hash codes and rigid clustering that fails to adapt to varying data densities, leading to suboptimal retrieval accuracy. To address these limitations, this paper proposes a novel framework called Dual-Stage Semantic Hashing with Adaptive Density-Aware Clustering (DSH-ADAC). The framework first uses a dual-branch CNN with attention-based fusion to generate binary hash codes that preserve both visual and contextual semantics. In the second stage, an adaptive density-aware DBSCAN algorithm is applied to cluster these hash codes, where the local density dynamically adjusts the clustering parameters to handle diverse data distributions better. The proposed method is utilized in intelligent image retrieval tasks, such as surveillance and e-commerce, providing efficient and meaningful image grouping and retrieval. Experimental results demonstrate an improvement of 96.8% in clustering quality and 97.3% in retrieval accuracy compared to existing approaches, showcasing the method's robustness by 98.2% and efficiency by 97.8% in real-world scenarios. © 2025 IEEE.

Intrusion detection in financial networks endures an essential challenge to preserving data safe and interrupting sophisticated cyber threats that target transactional systems. Cyber threats that target transactional systems are challenging to comprehend. Traditional machine learning and rule-based methods often fail to perform effectively because they are unable to recognize the complex, changing, and non-Euclidean structure of financial data or the intricate relationships between entities. This study presents a method for detecting intrusions in financial networks based on Graph Neural Networks (GNN). The technique leverages the capabilities of GNNs to model dependencies at the node, edge, and graph levels. The results of this study demonstrate that when current methods have not proven adequate, including their inability to depend on fixed attributes and evolve to deal with new forms of threats. It also offers a new architecture that takes into consideration node embeddings, temporal dynamics, and information about the graph's topology to ensure it can appropriately identify unusual behavior. The proposed model consistently outperforms traditional methods in terms of precision, recall, and F1 Score. The results can be demonstrated by experiments on datasets containing real-world financial transactions, and the significance of GNNs contributes to ensuring intrusion detection systems that are accurate, scalable, and adaptable. These outcomes have a significant impact on both the progress of fraud prevention and the real-time safety of financial systems. © 2025 IEEE.

Multi-object detection and tracking are a fundamental task in computer vision, especially in surveillance, autonomous systems, and robotics. YOLOv7, a state-of-the-art real-time object detection model, offers high-speed performance and accuracy, which faces challenges in maintaining object identity over time. Existing tracking methods often suffer from drift, occlusion, and identity switching due to noise, dynamic environments, and lack of temporal consistency. To address these issues, here is the proposed SMODT-Y7PF (Secure Multi-Object Detection and Tracking via YOLOv7 with Particle Filter), a hybrid framework integrating YOLOv7 for robust detection and Particle Filter algorithms for probabilistic tracking and motion estimation. This combination improves identity preservation and resilience against occlusion and noise. The proposed method uses YOLOv7 to detect multiple objects in each frame and employs Particle Filters to maintain object trajectories by estimating the posterior distribution over time. This integration ensures secure and consistent tracking in real-time applications. Experimental results show that SMODT-Y7PF significantly improves tracking accuracy, reduces identity switches, and enhances robustness in dynamic and cluttered environments, making it ideal for smart surveillance and autonomous navigation systems. © 2025 IEEE.

The increasing demand for e-learning emphasizes the degree of essential effective student support systems. This paper checks how artificial intelligence-fuel can be used as virtual teaching assistants to offer real-time educational assistance through natural language processing (NLP). This initiative creates a smart chatbot that is able to guide students in course materials, clarify ideas and provide feedback to questions. For other machine learning techniques comparatively, deep learning NLP models are considered to increase convergent accuracy and reference awareness. Using online students, the approach checks that this reaction reduces time, improves memoirs, and increases participation. According to the results of the experiments, the real-time support provided by the AI-based chatbots greatly improves student education experiences. Real-time tailored support provided by AI-based chatbots greatly improves student education experiences, according to results of experiments. Key findings suggest that incorporating virtual assistants driven by artificial intelligence into online learning will help to reduce the workload for human teachers and create a more interesting and easily available learning environment. This paper helps to develop smart e-learning systems and emphasizes the opportunities of artificial intelligence in improving digital education. Future studies will focus on enhancing chatbot adaptability, multilingual support, and emotional intelligence to better student contacts. © 2025 IEEE.