Robust ImageGraph: Rank-Level Feature Fusion for Image Search.

Area of Science:

• Computer Science
• Image Processing
• Machine Learning

Background:

• Feature fusion is effective for image search but susceptible to outliers (false positives) from poor features or parameters.
• Robustness to outliers is a critical challenge for fusion schemes in image retrieval systems.

Purpose of the Study:

• To propose a novel rank-level framework for robust feature fusion in image search.
• To mitigate the negative impact of outliers on image retrieval performance.

Main Methods:

• Defined Rank Distance to measure image relevance at the rank level.
• Introduced Bayes similarity to weight individual features, down-weighting outliers.
• Constructed a directed ImageGraph weighted by Bayes similarity to merge rank lists.
• Performed local ranking on the fused ImageGraph for outlier-robust re-ordering.

Main Results:

• The proposed rank-level fusion framework demonstrated effectiveness across four benchmark datasets.
• The method significantly outperformed two existing fusion schemes.
• Achieved competitive results compared to state-of-the-art image retrieval techniques.

Conclusions:

• The proposed rank-level framework offers a robust solution for feature fusion in image search.
• The approach effectively addresses the challenge of outliers, improving overall retrieval accuracy.
• The method provides a competitive and effective alternative to current fusion strategies.