Linear Subspace Ranking Hashing for Cross-Modal Retrieval

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Hashing is crucial for indexing and retrieving large-scale, high-dimensional multimedia data.
• Existing cross-modal hashing algorithms often use binary space partitioning functions.
• There is a need for more effective and flexible cross-modal hashing methods.

Purpose of the Study:

• To propose a novel ranking-based hashing framework for cross-modal retrieval.
• To map data from different modalities into a common Hamming space.
• To measure cross-modal similarity using Hamming distance.

Main Methods:

• Developed a novel hashing framework utilizing rank correlation measures.
• Jointly learned linear subspaces for each modality to preserve ranking orders.
• Employed a probabilistic approximation for efficient gradient descent optimization.
• Designed a flexible framework accommodating various loss functions.

Main Results:

• The proposed ranking-based hash functions are scale-invariant, numerically stable, and highly nonlinear.
• The framework efficiently solves the optimization problem using gradient descent.
• Achieved competitive performance against state-of-the-art methods on four real-world multimodal datasets.
• Demonstrated moderate training and testing times.

Conclusions:

• The novel ranking-based hashing framework offers an effective solution for cross-modal retrieval.
• The method provides flexibility and efficiency in handling high-dimensional multimedia data.
• This approach advances the field of cross-modal hashing for large-scale data indexing.