Discrete Latent Factor Model for Cross-Modal Hashing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Multimedia Systems

Background:

• Cross-modal hashing (CMH) is crucial for efficient similarity search across different data types.
• Existing CMH methods face a trade-off between training speed (continuous methods) and accuracy (discrete methods).

Purpose of the Study:

• To develop a novel CMH method that combines high accuracy with efficient training.
• To address the limitations of current relaxation-based continuous and discrete CMH approaches.

Main Methods:

• Propose Discrete Latent Factor model-based cross-modal Hashing (DLFH), a discrete CMH technique.
• DLFH directly learns binary hash codes for improved CMH performance.
• Focus on efficient training strategies for discrete hashing.

Main Results:

• DLFH achieves significantly better accuracy compared to existing CMH methods.
• The training time of DLFH is comparable to faster, less accurate relaxation-based continuous methods.
• DLFH outperforms traditional discrete methods in both accuracy and training efficiency.

Conclusions:

• DLFH offers a superior solution for cross-modal similarity search by balancing accuracy and speed.
• The proposed method effectively overcomes the accuracy-speed trade-off in existing CMH techniques.