Semantic-Aware Co-Indexing for Image Retrieval.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Content-based image retrieval (CBIR) relies on inverted indexes for efficient database access.
• Integrating diverse image features enhances retrieval perspectives but challenges efficiency and scalability.
• Current methods struggle with balancing retrieval performance and computational overhead.

Purpose of the Study:

• To propose a novel semantic-aware co-indexing algorithm for large-scale image retrieval.
• To enhance the discriminative power of inverted indexes by jointly embedding local features and semantic attributes.
• To ensure retrieval efficiency and scalability in content-based image retrieval systems.

Main Methods:

• Developed a semantic-aware co-indexing algorithm for large-scale image retrieval.
• Jointly embedded local invariant features and semantic attributes into inverted indexes.
• Utilized semantic attributes to refine initial local feature indexes by filtering and inserting semantically similar images.

Main Results:

• The proposed co-indexing method effectively enhances the discriminative capability of inverted indexes.
• Off-line co-indexing operations introduce minimal computational overhead to on-line retrieval.
• Experiments demonstrate competitive performance compared to recent retrieval methods.

Conclusions:

• The semantic-aware co-indexing algorithm offers a scalable and efficient solution for large-scale image retrieval.
• Combining local features with semantic attributes significantly improves retrieval accuracy.
• The method provides a distinct advantage over existing approaches by integrating cues during indexing rather than retrieval.