Object-guided multi-granularity unsupervised hashing for image retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised hashing methods for large-scale image retrieval face challenges with semantic similarity measurement.
• Existing methods often use cosine similarity, which is insufficient for capturing nuanced image relationships.
• Contrastive learning approaches struggle with unbalanced or inaccurate positive and negative examples.

Purpose of the Study:

• To propose a novel unsupervised hashing method, Object-guided Multi-granularity Unsupervised Hashing (OMUH).
• To enhance semantic similarity representation in image retrieval using object detection.
• To improve the performance of unsupervised hashing by refining example selection for contrastive learning.

Main Methods:

• Utilizing a pre-trained object detection model to generate pseudo-labels for images.
• Reconstructing the affinity matrix by emphasizing object roles based on pseudo-labels.
• Enlarging positive and purifying negative examples for contrastive learning using pseudo-labels.
• Designing a multi-granularity objective function to preserve intra-image, inter-image, and intra-class similarities.

Main Results:

• The proposed OMUH method demonstrates superior performance compared to state-of-the-art unsupervised hashing techniques.
• Experiments on three benchmark datasets show significant improvements, with mean Average Precision (mAP) increases up to 12%.
• The object-guided approach effectively enhances the semantic understanding for more accurate hash code learning.

Conclusions:

• OMUH offers a robust and effective solution for unsupervised hashing in large-scale image retrieval.
• Leveraging object detection provides a powerful mechanism to overcome limitations of traditional similarity measures.
• The method's ability to refine contrastive learning examples and preserve diverse similarities leads to substantial performance gains.