Flexible Cross-Modal Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Hashing is crucial for efficient large-scale data retrieval, offering low storage and high speed.
• Current cross-modal hashing methods rely on readily available sample correspondence, which is often unrealistic.
• Existing methods lack flexibility due to requirements for equal sample numbers across modalities.

Purpose of the Study:

• To propose a flexible cross-modal hashing approach (FlexCMH) for learning effective hashing codes from weakly paired data.
• To address the challenge of partially or totally unknown correspondence between samples across modalities.
• To overcome the inflexibility of existing methods regarding sample size disparities.

Main Methods:

• Introduced a clustering-based matching strategy to identify potential correspondence between samples across modalities.
• Jointly optimized potential correspondence, cross-modal hashing functions, and a hashing quantitative loss in a unified objective.
• Employed an alternative optimization technique to coordinate correspondence and hash functions for mutual reinforcement.

Main Results:

• FlexCMH demonstrated significantly improved performance compared to state-of-the-art methods on public multimodal datasets.
• The proposed method effectively handles weakly paired data with unknown or partial correspondence.
• Experiments confirmed FlexCMH's high degree of flexibility for practical cross-modal hashing applications.

Conclusions:

• FlexCMH provides a robust and flexible solution for cross-modal hashing, particularly with imperfectly paired data.
• The method's ability to learn from weakly paired data opens new possibilities for real-world multimodal retrieval systems.
• FlexCMH advances the field by enabling effective hashing despite challenges in data correspondence and sample distribution.