Adaptive Bit Selection for Scalable Deep Hashing.

Area of Science:

• Computer Science
• Artificial Intelligence

Background:

• Deep hashing is crucial for compact feature representation in content-based image retrieval.
• Existing methods often require training diverse models with varying resource demands.

Purpose of the Study:

• To propose a scalable deep hashing framework for generating binary codes of different lengths.
• To address the limitations of current deep hashing methods regarding computational costs and model diversity.

Main Methods:

• A novel framework employing iterative bit pool generation and adaptive bit selection using reinforcement learning.
• Optimization of retrieval performance and bit properties during training.
• Integration of existing binary hashing methods for scalable code generation.

Main Results:

• The proposed framework effectively generates binary codes with adaptive lengths.
• Experiments on four public datasets demonstrate superior performance in image retrieval tasks.
• The method proves effective in optimizing retrieval performance and bit properties.

Conclusions:

• The developed scalable deep hashing framework offers an efficient and flexible solution for image retrieval.
• The adaptive bit selection mechanism reduces computational overhead and enhances model adaptability.
• This approach provides a unified framework for various binary hashing methods.