Diffusion Parameters Analysis in a Content-Based Image Retrieval Task for Mobile Vision.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Image retrieval tasks increasingly leverage feature distribution for enhanced model performance.
• Diffusion algorithms, combining manifold data and k-Nearest Neighbors (kNN) graphs, are common for this purpose.
• Mobile applications face computational constraints for large-scale kNN graph construction.

Purpose of the Study:

• Optimize diffusion for image retrieval in mobile vision applications.
• Achieve a balance between computational efficiency and retrieval performance.
• Address the computational infeasibility of exhaustive kNN graph creation on mobile devices.

Main Methods:

• Employed approximate algorithms for kNN graph construction to reduce computational complexity.
• Utilized a genetic algorithm (GA) for optimizing diffusion parameters.
• Validated the approach on public datasets including Oxford5k, ROxford5k, Paris6k, RParis6k, and Oxford105k.

Main Results:

• The genetic algorithm effectively identified optimal diffusion parameters, surpassing grid search in efficiency and effectiveness.
• The proposed method achieved high retrieval performance despite using approximate kNN graphs.
• Demonstrated the GA's capability for global search in the diffusion parameter space, which is unfeasible with exhaustive methods.

Conclusions:

• Optimized diffusion using a genetic algorithm provides a computationally efficient yet high-performing solution for mobile image retrieval.
• Genetic algorithms offer a superior approach to parameter optimization compared to traditional methods like grid search for diffusion algorithms.
• The method successfully balances computational load and retrieval accuracy for practical mobile vision applications.