Deep Discrete Supervised Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Hashing is crucial for large-scale search, offering low storage and fast queries.
• Supervised hashing methods outperform unsupervised ones by leveraging labeled data.
• Recent advances include discrete supervised hashing and deep hashing for improved performance.

Purpose of the Study:

• To develop a novel deep hashing method that integrates discrete coding and deep feature learning.
• To enable direct guidance of both procedures using supervised information within a unified framework.
• To enhance the feedback loop between feature learning and hash-code generation for superior image retrieval.

Main Methods:

• Proposes Deep Discrete Supervised Hashing (DDSH), a novel deep hashing framework.
• Utilizes pairwise supervised information to directly guide discrete coding and deep feature learning.
• Integrates these two procedures to create a synergistic feedback mechanism.

Main Results:

• DDSH demonstrates superior performance compared to state-of-the-art discrete and deep hashing methods.
• Experiments conducted on four real-world datasets validate the effectiveness of DDSH.
• The method achieves significant improvements in image retrieval tasks.

Conclusions:

• DDSH is the first deep hashing method to jointly guide discrete coding and deep feature learning.
• The proposed approach effectively enhances the interplay between feature learning and hash-code generation.
• DDSH offers a promising solution for accurate and efficient large-scale image retrieval.