Coupled binary embedding for large-scale image retrieval.

Area of Science:

• Computer Vision
• Information Retrieval
• Machine Learning

Background:

• Visual matching is key in bag-of-words (BoW) image retrieval.
• SIFT descriptors lose discriminative power during quantization and only capture local texture.
• These limitations impair BoW model performance and increase false positive matches.

Purpose of the Study:

• To improve the discriminative power of visual matching in BoW models.
• To reduce false positive matches in image retrieval.
• To enhance the overall precision and efficiency of image retrieval systems.

Main Methods:

• Embedding multiple binary features at the indexing level.
• Utilizing a multi-IDF scheme to model feature correlations and couple binary features into the inverted file.
• Incorporating binary feature matching verification methods like Hamming embedding.
• Exploring the fusion of binary color features and global color features.

Main Results:

• The proposed method significantly enhances the precision of visual matching by integrating SIFT and binary features.
• False positive matches are substantially reduced.
• Extensive experiments on benchmark datasets demonstrate significant improvements over the baseline approach.
• Large-scale tests show acceptable memory usage and query times.

Conclusions:

• The integration of multiple binary features, coupled with a multi-IDF scheme, effectively addresses the limitations of SIFT descriptors in BoW image retrieval.
• The framework allows for the seamless incorporation of binary feature matching verification and color feature fusion.
• The proposed approach offers a significant improvement in visual matching precision and overall retrieval performance, achieving competitive results with state-of-the-art methods.