A Novel Deep Learning-Based Framework for Semantic Image Retrieval Using Multi-Scale Attention and Contrastive Learning

Semantic image retrieval plays a vital role in various domains, including healthcare, e-commerce, and multimedia systems, where accurate and efficient retrieval of semantically relevant images is essential. This paper presents a novel deep learning-based framework for semantic image retrieval that integrates multi-scale attention mechanisms and hybrid contrastive learning to address critical limitations of existing methods. The proposed framework dynamically captures global and local features across varying scales, ensuring robustness to spatial and contextual variations, while the hybrid contrastive learning strategy enhances inter-class separability and intra-class compactness in the embedding space. The framework's performance was evaluated on three benchmark datasets—ImageNet, Oxford5k, and Paris6k—utilizing key metrics such as mean average precision (mAP), recall, precision, F1-score, and retrieval time. The proposed method consistently demonstrated high accuracy and computational efficiency, with mAP scores of 0.903, 0.891, and 0.879 on ImageNet, Oxford5k, and Paris6k, respectively, and an average retrieval time of 9.65 ms per query. These results indicate the framework's capacity to deliver robust and scalable performance across diverse datasets. This study underscores the potential of the proposed framework for large-scale and real-time semantic image retrieval applications. By addressing challenges in fine-grained retrieval and scalability, this work establishes a new benchmark in the field and elucidates pathways for future advancements in multimodal and unsupervised learning approaches.