Sub-Region Localized Hashing for Fine-Grained Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained image hashing faces challenges in extracting discriminative local information.
• Existing dense attention mechanisms struggle with diverse local features and intra-class variations.

Purpose of the Study:

• To propose a novel sub-Region Localized Hashing (sRLH) method.
• To learn intra-class compact and inter-class separable hash codes for efficient fine-grained image retrieval.

Main Methods:

• Developed a sub-region localization module to identify discriminative local features by locating peaks in feature maps.
• Employed Gram-Schmidt orthogonalization on binary centers for inter-class separability.
• Enforced hash codes of the same class to converge to a common binary center.

Main Results:

• The proposed sRLH method demonstrates superior performance compared to state-of-the-art methods.
• Achieved improved fine-grained image retrieval accuracy on four benchmark datasets.
• Successfully captured diverse subtle local information and mitigated intra-class variations.

Conclusions:

• sRLH effectively addresses the limitations of existing fine-grained image hashing techniques.
• The method offers a promising approach for efficient and accurate fine-grained image retrieval.
• Source code will be publicly released to facilitate further research.