Content-Based Image Retrieva‎l (CBIR) is a technique that identifies similar images within large archives. This approach can be beneficial in various domains, including medicine, forensics, localization, and more. The field of CBIR has rapidly evolved, garnering significant attention from researchers over the past two decades. In this study, a novel CBIR system is proposed that enhances performance by integrating advanced artificial intelligence techniques such as attention mechanisms, Graph Neural Networks (GNNs), and multimodal learning models like BLIP. The proposed system employs multimodal models and integrates them with graph neural networks to perform feature extraction in a novel manner. It utilizes high-dimensional tensors as inputs to a GNN architecture, which models the relationships among different components of an image. Attention mechanisms are leveraged to focus on the most relevant parts of the image, while GNNs capture the structural and contextual dependencies between image elements, enabling the system to deliver more accurate results. Furthermore, the BLIP multimodal model is utilized to expand the system's capabilities, facilitating the connection between visual and textual data for targeted identification of the most significant image components. This multimodal approach enables a deeper understanding of the relationship between images and their descriptive text, significantly enhancing retrieva‎l effectiveness for visual-textual hybrid data. This research, by demonstrating the potential of combining attention mechanisms, GNNs, and multimodal learning for a CBIR system, introduces a new approach in this field. Experimental eva‎luations indicate that the proposed approach achieves an accuracy of 85.67% in retrieving the most similar image to the query image.

محمدرضا احمدزاده

الهام محمودزاده , محمد داورپناه جزي