Discrete Semantic Alignment Hashing for Cross-Media Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multimodal data (e.g., images, text) necessitates cross-media hashing for efficient retrieval.
• Existing methods often overlook semantic gaps between modalities, limiting retrieval effectiveness.
• Image features lack semantics compared to text keywords, creating an unbalanced semantic gap.

Purpose of the Study:

• To propose a novel cross-media hashing method, discrete semantic alignment hashing (DSAH), to address unbalanced semantic gaps.
• To enhance hash code quality and retrieval performance in multimodal datasets.
• To develop a method that seamlessly incorporates attributes for improved alignment.

Main Methods:

• DSAH utilizes collaborative filtering to mine class label-hash code relations, reducing computational costs.
• Image attributes are employed to align semantic information between image and text modalities.
• A discrete optimization algorithm directly learns discrete hash codes with closed-form solutions for each bit.

Main Results:

• The proposed DSAH model demonstrates effective incorporation of attributes for cross-media retrieval.
• Experiments on public databases show promising retrieval performance compared to existing methods.
• The method successfully reduces memory consumption and computational cost.

Conclusions:

• DSAH offers a simple yet effective solution for cross-media retrieval challenges posed by semantic gaps.
• The discrete optimization approach yields high-quality hash codes.
• This method advances the field of cross-media hashing by improving retrieval accuracy and efficiency.