Label Consistent Matrix Factorization Hashing for Large-Scale Cross-Modal Similarity Search.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multimodal hashing is crucial for efficient cross-modal similarity search in large datasets.
• Supervised methods enhance accuracy by using data labels, but often struggle with scalability and information loss.
• Existing methods transform labels into pairwise similarities, leading to storage issues and reduced discriminative power.

Purpose of the Study:

• To develop a scalable and effective supervised multimodal hashing method that directly utilizes semantic labels.
• To improve the discriminative power of hash codes for enhanced cross-modal retrieval accuracy.
• To address the limitations of pairwise similarity transformations in current supervised hashing techniques.

Main Methods:

• Introduced Label Consistent Matrix Factorization Hashing (LCMFH), a novel supervised multimodal hashing algorithm.
• Transformed heterogeneous data into latent semantic spaces, ensuring same-category data share representations.
• Directly leveraged semantic labels to guide the hashing learning process, avoiding pairwise similarity computations.

Main Results:

• LCMFH demonstrated superior performance compared to several state-of-the-art methods in experiments.
• The proposed method effectively preserves category information from semantic labels.
• Hash codes generated by LCMFH exhibit enhanced discriminative power for cross-modal similarity search.

Conclusions:

• LCMFH offers a simple yet effective approach for supervised multimodal hashing.
• Directly utilizing semantic labels leads to more accurate and scalable cross-modal similarity search.
• The method's ability to learn discriminative hash codes makes it suitable for large-scale multimedia analysis.