Cross-modal hashing with missing labels.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hashing-based cross-modal retrieval offers storage and speed benefits.
• Existing methods often assume complete labels, which is unrealistic.
• Missing labels and ignored label similarities are key challenges in real-world applications.

Purpose of the Study:

• To develop a novel method for cross-modal retrieval that addresses missing labels and label correlations.
• To improve the accuracy and robustness of hashing-based cross-modal retrieval systems.

Main Methods:

• Proposed Cross-Modal Hashing with Missing Labels (CMHML) method.
• Introduced Reliable Label Learning to utilize observed labels.
• Decomposed labels into latent representations for inferring missing information.
• Incorporated Label Correlation Preservation to capture semantic relationships.
• Employed Global Approximation Learning for hash code generation.
• Constructed a similarity matrix from predicted labels to guide hash code learning.
• Trained linear classifiers for mapping samples to a low-dimensional Hamming space.

Main Results:

• CMHML demonstrates competitive performance against state-of-the-art methods.
• The model is effective even when a significant portion of labels is missing.
• Extensive experiments on four public datasets validate the proposed approach.

Conclusions:

• CMHML successfully addresses the limitations of existing cross-modal hashing methods.
• The proposed techniques for handling missing labels and label correlations enhance retrieval performance.
• CMHML offers a robust solution for practical cross-modal retrieval scenarios with imperfect data.