Relational Consistency Induced Self-Supervised Hashing for Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Large-scale image retrieval demands efficient methods for searching vast datasets.
• Existing hashing techniques often struggle to capture complex semantic relationships within image data.
• Self-supervised learning offers a promising avenue for learning representations without manual annotations.

Purpose of the Study:

• To propose a novel hashing framework, relational consistency induced self-supervised hashing (RCSH), for large-scale image retrieval.
• To effectively capture and preserve semantic data relationships in both latent and Hamming spaces.
• To improve the accuracy and generalization ability of image retrieval systems.

Main Methods:

• RCSH explores relational consistency between data samples across different feature spaces.
• It learns data relationships by identifying prototypes that group similar images in the latent space.
• A dual prototype contrastive loss is introduced to align prototype assignments between latent and Hamming spaces, ensuring consistent data-to-prototype and data-to-data relationships.

Main Results:

• The proposed RCSH method significantly outperforms state-of-the-art methods on four benchmark image retrieval datasets.
• RCSH demonstrates strong performance in out-of-domain retrieval tasks, indicating excellent generalization capabilities.
• The framework successfully learns reliable data relationships and preserves them in compact hash codes.

Conclusions:

• RCSH provides an effective self-supervised hashing approach for large-scale image retrieval.
• The method's ability to capture semantic structures enhances retrieval accuracy and robustness.
• The framework's generalization ability makes it suitable for diverse real-world applications.