Discriminative Geometric-Structure-Based Deep Hashing for Large-Scale Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep hashing is a mainstream technique for large-scale image retrieval, encoding images into hash codes while preserving feature similarity.
• Existing methods focus on geometric-structure preservation but inadequately ensure feature discrimination, which is crucial for retrieval performance.

Purpose of the Study:

• To propose a novel discriminative geometric-structure-based deep hashing (DGDH) method to improve feature discrimination for enhanced image retrieval.
• To introduce three new loss terms based on class centers to induce a discriminative geometrical structure.

Main Methods:

• The DGDH method incorporates a margin-aware center loss for intraclass compactness, a linear classifier for interclass separability, and a radius loss to reduce quantization errors.
• An efficient alternating optimization algorithm with guaranteed convergence is used to optimize the DGDH model.
• Theoretical analysis of the method's robustness and generalization is provided.

Main Results:

• The proposed DGDH method significantly improves feature discrimination compared to existing approaches.
• Experiments on five benchmark datasets demonstrate superior image retrieval performance of DGDH over state-of-the-art methods.
• The developed loss terms effectively enhance both intraclass compactness and interclass separability.

Conclusions:

• Improving feature discrimination is essential for advancing deep hashing-based image retrieval.
• The proposed DGDH method effectively addresses the limitations of existing techniques by introducing a discriminative geometrical structure.
• DGDH offers a promising direction for future research in large-scale image retrieval.