Content-Based Image Retrieval using Support Vector Machine and Clustering
Subject Areas : International Journal of Data Envelopment Analysis
محمدمهدی علیان نژادی
1
*
,
Ahmad Kamandi
2
,
Fatemeh Ahangar Darabi
3
,
Hesamoddin Pourrostami
4
,
سید محمدرضا هاشمی
5
1 - گروه علوم کامپیوتر، دانشگاه علم و فناوری مازندران، بهشهر، ایران
2 - Department of Mathematics, University of Science and Technology of Mazandaran, Iran
3 - Department of Mathematics, University of Science and Technology of Mazandaran, Behshahr, Iran
4 - Department of Information Management, International Graduate School of Artificial Intelligence, National Yunlin University of Science and Technology, Douliu, Taiwan
5 - گروه مهندسی کامپیوتر و فناوری اطلاعات، دانشگاه صنعتی شاهرود، شاهرود، ایران.
Keywords:
Abstract :
Accurate content-based image retrieval (CBIR) is critical for applications ranging from large-scale digital libraries to real-time mobile search, yet many high-precision methods suffer from prohibitive query-time costs as database sizes grow. To address this challenge, we introduce a scalable two-stage retrieval framework that balances accuracy with efficiency. First, we extract deep convolutional features from all images using a pre-trained ResNet50 model and partition the feature space into semantically coherent clusters via K-means. During retrieval, an incoming query image undergoes feature extraction and is quickly assigned to the most relevant cluster using a directed acyclic graph support vector machine classifier. Within the selected cluster, we perform fine-grained similarity ranking to produce the final set of retrieved images. We validate our approach on the Corel-10k dataset, demonstrating that our method achieves a recall of 0.91 and a precision of 0.93, while reducing average query time by 45% compared to a non-clustered baseline. Ablation studies reveal the impact of cluster granularity and classifier choice on retrieval performance. The proposed framework offers an effective solution for large-scale CBIR, maintaining high retrieval precision with substantially lower latency.