MTFH: A Matrix Tri-Factorization Hashing Framework for Efficient Cross-Modal Retrieval.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Hashing significantly enhances cross-modal retrieval due to low storage and high query speeds.
• Existing methods often use unified or equal-length hash codes, limiting scalability for data with non-corresponding modalities.
• This limitation hinders efficient encoding and comparability of heterogeneous data.

Purpose of the Study:

• To address the limitations of fixed-length hash codes in cross-modal retrieval.
• To propose a flexible framework capable of encoding heterogeneous data with varying hash lengths.
• To generalize cross-modal retrieval across diverse and challenging scenarios, including paired/unpaired data.

Main Methods:

• Introduced Matrix Tri-Factorization Hashing (MTFH), a generalized and flexible cross-modal hashing framework.
• Developed an efficient objective function to learn modality-specific hash codes with flexible length settings.
• Synchronously learned two semantic correlation matrices to ensure comparability of heterogeneous data representations.

Main Results:

• MTFH effectively handles paired or unpaired multi-modal data and varying hash length encoding.
• The framework generates more semantically meaningful hash codes for challenging cross-modal retrieval tasks.
• Experiments on benchmark datasets demonstrate MTFH's superiority and competitive performance.

Conclusions:

• MTFH offers a superior and flexible approach to cross-modal hashing and retrieval.
• The method enhances representation scalability and semantic correlation for heterogeneous data.
• MTFH achieves state-of-the-art performance across various challenging retrieval scenarios.