A Novel Adaptive Feature Fusion Strategy for Image Retrieval

Area of Science:

• Computer Science
• Information Retrieval
• Artificial Intelligence

Background:

• Efficient image retrieval is crucial in the big data era.
• Single-feature retrieval systems have limitations.
• Multi-feature fusion enhances image retrieval performance.

Purpose of the Study:

• To propose an adaptive multi-feature fusion strategy for content-based image retrieval.
• To improve retrieval accuracy and generalization by dynamically selecting and weighting features.

Main Methods:

• Image feature extraction and initial similarity calculation using information entropy.
• Automatic effective feature selection based on retrieval trust derived from single-feature precision.
• Feature weight optimization using the PageRank algorithm.
• Comprehensive similarity calculation for final retrieval results.

Main Results:

• The proposed method achieved high top-10 retrieval precision: 99.55% (Corel1k), 88.02% (UC Merced Land-Use), and 88.28% (RSSCN7).
• The mean average precision (mAP) on the Holidays dataset was 92.46%.
• The adaptive fusion strategy outperformed single-feature and fixed-feature fusion methods.

Conclusions:

• The adaptive multi-feature fusion strategy offers superior performance and generalization compared to single-feature methods.
• Dynamic feature selection and fusion optimize retrieval effectiveness for each query.
• The method effectively addresses the challenges of image retrieval in large datasets.