Joint Hypergraph Learning for Tag-Based Image Retrieval.

Area of Science:

• Computer Science
• Information Retrieval
• Multimedia Systems

Background:

• Image sharing platforms like Flickr are increasingly popular, driving research in tag-based image retrieval.
• Existing methods often process visual features and tags separately, limiting retrieval effectiveness.
• Social user-contributed images require efficient retrieval methods leveraging rich metadata and visual content.

Purpose of the Study:

• To propose a novel approach for tag-based image retrieval that effectively fuses global and local visual features.
• To enhance image relevance learning by integrating tag information within a hypergraph framework.
• To improve the accuracy of image retrieval by utilizing a pseudo-relevance feedback mechanism.

Main Methods:

• Constructing a hypergraph incorporating global visual features, local visual features, and tag information.
• Implementing a pseudo-relevance feedback mechanism to identify potentially relevant images.
• Applying hypergraph learning algorithms to compute image relevance scores for query matching.

Main Results:

• The proposed hypergraph approach demonstrates superior performance in tag-based image retrieval.
• Fusion of global and local visual features significantly improves image relevance compared to separate usage.
• The pseudo-relevance feedback mechanism effectively refines retrieval accuracy.

Conclusions:

• The developed hypergraph learning method offers an effective solution for tag-based image retrieval.
• Feature fusion and pseudo-relevance feedback are crucial for enhancing the relevance of retrieved images.
• This approach provides a robust framework for searching images contributed by social users.