Learning Discriminative Binary Codes for Large-scale Cross-modal Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Cross-modal retrieval leverages heterogeneous data by learning shared representations.
• Hashing methods aim to generate compact binary codes for efficient data indexing and search.
• Current methods often relax discrete constraints, leading to suboptimal binary codes.

Purpose of the Study:

• To develop a novel cross-modal hashing method that directly learns discriminative binary codes.
• To address the limitations of existing methods in exploring class label discriminative properties and preserving discrete constraints.
• To improve the accuracy and discriminative power of binary codes for cross-modal retrieval.

Main Methods:

• Proposed Discrete Cross-Modal Hashing (DCH) method.
• Learning modality-specific hash functions for unified binary code generation.
• Developing a discrete optimization algorithm to jointly learn hash functions and binary codes.

Main Results:

• DCH successfully learns discriminative binary codes while preserving discrete constraints.
• The proposed method outperforms existing approaches in cross-modal retrieval tasks.
• Experiments on benchmark datasets demonstrate state-of-the-art performance.

Conclusions:

• Discrete Cross-Modal Hashing (DCH) offers a superior approach to cross-modal retrieval.
• Directly learning discrete and discriminative binary codes enhances retrieval performance.
• The developed optimization algorithm effectively handles the joint learning problem.