Exploring Hierarchical Information in Hyperbolic Space for Self-Supervised Image Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Real-world image datasets exhibit inherent hierarchical structures.
• Existing self-supervised hashing methods fail to leverage this hierarchical information.
• This leads to suboptimal preservation of data relationships in learned hash codes.

Purpose of the Study:

• To apply visual hierarchical information to self-supervised hash learning.
• To address challenges in constructing, embedding, and exploiting visual hierarchies.
• To propose a novel method, Hierarchical Hyperbolic Contrastive Hashing (HHCH).

Main Methods:

• Embedding continuous hash codes into hyperbolic space for reduced distortion.
• Updating K-Means for adaptive construction of hierarchical semantic structures in hyperbolic space.
• Developing hierarchical contrastive learning (instance-wise and prototype-wise) to exploit these structures.

Main Results:

• HHCH demonstrates superior performance compared to state-of-the-art self-supervised hashing methods.
• Experiments conducted on four benchmark datasets validate the proposed approach.
• The method effectively captures and utilizes visual hierarchies for improved hashing.

Conclusions:

• The proposed HHCH method effectively leverages visual hierarchies in hyperbolic space for self-supervised hashing.
• This approach enhances the accuracy of learned hash codes and image retrieval performance.
• HHCH offers a significant advancement in self-supervised learning for large-scale image datasets.