Generalized Semantic Preserving Hashing for Cross-Modal Retrieval

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Multimedia data is rapidly increasing, necessitating efficient retrieval methods.
• Hashing techniques offer a scalable solution for large-scale cross-modal retrieval.
• Existing methods face challenges with diverse label associations and data correspondences.

Purpose of the Study:

• To propose a unified hashing framework for cross-modal retrieval adaptable to various data scenarios.
• To effectively capture semantic relationships between data items from different modalities.
• To address challenges posed by single-label, multi-label, and varied correspondence data.

Main Methods:

• A two-stage hashing framework is proposed.
• Stage one: Factorizing an affinity matrix to learn optimal hash codes.
• Stage two: Employing ridge regression and kernel logistic regression to learn hash functions for mapping data to the bit domain.

Main Results:

• The framework demonstrates effectiveness across single-label (Wiki) and multi-label (MirFlickr, NUS-WIDE, Pascal, LabelMe) datasets.
• The proposed approach shows competitive performance compared to state-of-the-art methods.
• A novel iterative solution is presented for handling very large or incomplete training data.

Conclusions:

• The proposed hashing framework is versatile and effective for cross-modal retrieval tasks.
• It successfully captures semantic relationships and handles diverse data labeling scenarios.
• The iterative solution enhances scalability for large-scale machine learning applications.