Discrete Two-Step Cross-Modal Hashing through the Exploitation of Pairwise Relations

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Cross-modal hashing enables mapping heterogeneous data to binary codes for semantic similarity.
• Existing supervised methods often underutilize supervised information and suffer from unstable learning due to one-directional mapping.

Purpose of the Study:

• To propose a novel supervised cross-modal hashing method, Discrete Two-step Cross-modal Hashing (DTCH).
• To address limitations of existing methods by fully exploiting pairwise relations and stabilizing the hash learning process.

Main Methods:

• DTCH utilizes pairwise similarity relations from supervision information.
• Combines matrix factorization and label regression for label matrix, and a semirelaxed/semidiscrete strategy for pairwise similarity matrix.
• Incorporates fine-grained features and an out-of-sample extension strategy.

Main Results:

• The proposed DTCH method demonstrates superior performance in cross-modal retrieval tasks.
• Experimental validation on two widely used datasets confirms the method's effectiveness.
• DTCH reduces cumulative quantization errors and enhances retrieval efficiency and accuracy.

Conclusions:

• DTCH offers a more stable and effective approach to supervised cross-modal hashing.
• The method successfully preserves consistency between different modal distributions and pairwise similarity relations.
• DTCH advances the field of cross-modal retrieval through its innovative techniques.