Multi-relationship based relevance feedback scheme in web image retrieval

作者：Jin Hai^*; He Ruhan; Tao Wenbing

来源：International Journal of Innovative Computing Information and Control, 2008, 4(6): 1315-1324.

• web image retrieval
• Relevance Feedback (RF)
• Manifold Ranking Algorithm (MRA)
• Similarity Propagation Algorithm (SPA)

Traditional Relevance Feedback (RF) in web image retrieval usually combines the two modalities of web images, i.e. text and visual content, to improve the retrieval performance. However, the combination scheme is either sequential or linear, and the used similarity metric is simply based on pair-wise distance. Therefore, the multiple relationships within and between the two modalities of web images are not fully explored. In this paper, a novel RF scheme, named Multi-Relationship Based Relevance Feedback (MRBRF), is proposed, which simultaneously utilizes both the intra-modal relationship and the inter-modal relationship among web images at each iteration of RF. The intra-modal relationship reveals the intrinsic global manifold structure within the textual and the visual feature spaces of web images, and the inter-modal relationship reflects the relationship between the two feature spaces via web hyperlinks. In this new approach, Manifold Ranking Algorithm (MRA) and Similarity Propagation Algorithm (SPA) are integrated to explore the multiple relationships of web images. The experiments are carried out in our web image retrieval system, named VAST (VisuAl & SemanTic image search), and the results show the effectiveness of the proposed MRBRF scheme.

• 出版日期2008-6
• 单位华中科技大学