Unsupervised Topic Hypergraph Hashing for Efficient Mobile Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hashing is crucial for efficient mobile image retrieval, offering low data transmission and fast response times.
• Existing hashing methods often rely on low-level features, limiting discriminative capability and failing to capture high-order semantic correlations in images.

Purpose of the Study:

• To propose a novel unsupervised hashing scheme, Topic Hypergraph Hashing (THH), to address limitations in current image retrieval techniques.
• To enhance hashing codes by incorporating auxiliary text and modeling complex semantic relationships between images.

Main Methods:

• Discovered image-topic relations using robust collective non-negative matrix factorization.
• Constructed a unified topic hypergraph representing images and topics to model high-order semantic correlations.
• Learned hashing codes and functions by enforcing semantic consistency and preserving discovered semantic relations.

Main Results:

• THH effectively mitigates semantic shortages in hashing codes by utilizing auxiliary texts.
• The proposed method successfully models high-order semantic correlations among images.
• Experiments demonstrated superior performance of THH over state-of-the-art methods in mobile image retrieval.

Conclusions:

• Topic Hypergraph Hashing (THH) offers a significant advancement in unsupervised hashing for image retrieval.
• THH's ability to exploit semantic information makes it highly suitable for efficient mobile image retrieval applications.