Object-Level Visual-Text Correlation Graph Hashing for Unsupervised Cross-Modal Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Cross-modal hashing maps high-dimensional data to binary codes for efficient retrieval using Hamming distance.
• Unsupervised cross-modal hashing leverages paired data, making it suitable for real-world applications.
• Existing methods often overlook intra-modality and inter-modality correlations, limiting retrieval accuracy.

Purpose of the Study:

• To propose a novel approach, Visual-textful Correlation Graph Hashing (OVCGH), to address limitations in unsupervised cross-modal hashing.
• To enhance cross-modal retrieval accuracy by mining fine-grained object-level similarities.
• To effectively consider both intra-modality and inter-modality correlations within data.

Main Methods:

• Developed an unsupervised intra-modality correlation graph to learn representations of image regions and tags, capturing intra-modal dependencies.
• Designed a visual-text dependency building module to model relationships between image objects and text tags, capturing inter-modal correlations.
• Utilized object-level similarity mining to improve cross-modal hashing performance while suppressing noise.

Main Results:

• The proposed OVCGH approach effectively mines fine-grained object-level similarities in cross-modal data.
• The method demonstrates improved retrieval accuracy by considering both intra-modality and inter-modality correlations.
• Experimental validation on two datasets confirms the effectiveness of the OVCGH approach.

Conclusions:

• Object-level correlation modeling is crucial for enhancing unsupervised cross-modal hashing.
• The OVCGH approach provides a robust framework for capturing complex relationships in multi-modal data.
• This work advances the field of cross-modal retrieval through improved accuracy and applicability.