Content-Based Image Retrieval Using Transfer Learning and Vector Database

Abstract

Content-based image retrieval (CBIR) systems are essential for efficiently searching large image datasets using image features instead of text annotations. Major challenges include extracting effective feature representations to improve accuracy, as well as indexing them to improve the retrieval speed. The use of pre-trained deep learning models to extract features has elicited interest from researchers. In addition, the emergence of open-source vector databases allows efficient vector indexing which significantly increases the speed of similarity search. This paper introduces a novel CBIR system that combines transfer learning with vector databases to improve retrieval speed and accuracy. Using a pre-trained VGG-16 model, we extract high-dimensional feature vectors from images, which are stored and retrieved using the Milvus vector database. Our approach significantly reduces retrieval time, achieving real-time responses while maintaining high precision and recall. Experiments conducted on ImageClef, ImageNet, and Corel-1k datasets demonstrate the system’s effectiveness in large-scale image retrieval tasks, outperforming traditional methods in both speed and accuracy.