Unifying Sum and Weighted Aggregations for Efficient yet Effective Image Representation Computation.

Area of Science:

• Computer Science
• Image Processing
• Machine Learning

Background:

• Image search commonly uses aggregated local descriptors (e.g., SIFT) for vector representation.
• Standard sum aggregation is efficient but lacks discriminative power.
• Weighted aggregation offers better retrieval but incurs high computational costs.

Purpose of the Study:

• To develop a general mixed aggregation method unifying sum and weighted approaches.
• To balance the trade-off between retrieval quality and image representation efficiency.
• To enhance query performance through optimized weighting coefficient computation.

Main Methods:

• Introduced a general mixed aggregation framework for image representation.
• Proposed partitioning aggregated vectors into components to compute multiple weighting coefficients.
• Evaluated the method on standard public image retrieval benchmarks.

Main Results:

• The proposed mixed aggregation method achieves state-of-the-art performance in image retrieval.
• Demonstrated over ten times speedup compared to existing baseline methods.
• The general formulation effectively balances retrieval quality and representation efficiency.

Conclusions:

• The novel mixed aggregation method offers a superior solution for image search.
• The technique provides significant improvements in both accuracy and computational efficiency.
• This approach advances the field of image retrieval by optimizing descriptor aggregation.