A Probabilistic Approach to Cross-Region Matching-Based Image Retrieval

Area of Science:

• Computer Vision
• Machine Learning
• Information Retrieval

Background:

• Cross-Region Matching (CRM) excels in image retrieval by comparing image regions, offering robustness to geometric variations.
• CRM's performance is limited by its failure to model the distribution of visual concepts within image regions, reducing comparison precision.

Purpose of the Study:

• To provide a rigorous probabilistic interpretation of CRM-based image retrieval.
• To address the fundamental issue of ignoring visual concept distributions in CRM.
• To propose an improved image retrieval method utilizing deep convolutional features.

Main Methods:

• Scrutinized CRM-based image retrieval through a probabilistic interpretation following the probability ranking principle.
• Developed a novel approach treating locally clustered image regions as pseudo-labeled classes to model visual concept distributions.
• Implemented both non-parametric and parametric methods for modeling distributions, with probabilistic justification.

Main Results:

• The proposed pseudo-label approach significantly outperforms CRM and other comparable methods in image retrieval.
• Demonstrated superior performance on multiple benchmark datasets, with improvements exceeding 10 percentage points over CRM.
• The method effectively models visual concept distributions, leading to more precise similarity comparisons.

Conclusions:

• The pseudo-label approach effectively models visual concept distributions, enhancing image retrieval precision.
• This novel method represents a significant advancement over existing CRM techniques for robust image retrieval.
• The findings highlight the importance of modeling visual concept distributions for accurate image similarity assessment.