Semantic Neighbor Graph Hashing for Multimodal Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Hashing methods are crucial for efficient approximate nearest neighbor search.
• Existing multimodal hashing methods often overlook intra-class and inter-class variations.
• Preserving fine-grained similarity in multimodal data is challenging.

Purpose of the Study:

• To propose a novel multimodal hashing method, Semantic Neighbor Graph Hashing (SNGH).
• To enhance approximate nearest neighbor search by capturing detailed similarity metrics.
• To address limitations of existing methods in handling variations within and between classes.

Main Methods:

• Constructing a semantic graph incorporating semantic supervision and local neighborhood structure.
• Developing a function to adaptively calculate multi-level similarities, encoding variations.
• Employing logistic regression with kernel trick for modality-specific hash function learning.

Main Results:

• SNGH demonstrates superior performance over state-of-the-art methods.
• Experimental results validate the effectiveness on four benchmark datasets.
• The method successfully preserves fine-grained similarity metrics.

Conclusions:

• SNGH offers a significant advancement in multimodal hashing.
• The proposed approach effectively handles intra-class and inter-class variations.
• SNGH provides a robust solution for approximate nearest neighbor search in multimodal data.