Unsupervised Dual Deep Hashing With Semantic-Index and Content-Code for Cross-Modal Retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hashing technology offers efficient cross-modal retrieval but faces challenges with large datasets and semantic gaps.
• Supervised methods require extensive annotations, while unsupervised methods struggle with computational costs and preserving semantic information.

Purpose of the Study:

• To propose an unsupervised dual deep hashing (UDDH) method for effective cross-modal retrieval.
• To address the limitations of existing supervised and unsupervised methods in handling large-scale data and semantic discrepancies.

Main Methods:

• UDDH utilizes deep hashing networks to extract features and generate dual hashing codes (head code on semantic index, tail codes on modality content).
• A common semantic index and modality content codes are jointly learned to bridge semantic and heterogeneous gaps.
• The model integrates deep feature extraction, binary optimization, semantic index learning, and content code generation within a unified framework.

Main Results:

• UDDH significantly shrinks the search space by enabling queries to search within the same semantic index.
• The proposed method demonstrates superior retrieval efficiency and accuracy compared to state-of-the-art baselines.
• Collaborative optimization within the unified model enhances overall cross-modal retrieval performance.

Conclusions:

• UDDH offers an effective unsupervised approach for cross-modal retrieval, overcoming limitations of existing methods.
• The dual hashing architecture with semantic indexing provides a scalable and efficient solution for large-scale retrieval tasks.
• The integrated learning framework optimizes feature extraction and hashing for improved cross-modal understanding.