Concept Preserving Hashing for Semantic Image Retrieval With Concept Drift

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current hashing-based image retrieval assumes static databases, failing with dynamic updates and concept drift.
• Online hashing methods exist but do not address concept drift and are computationally expensive due to re-hashing all data.

Purpose of the Study:

• To propose Concept Preserving Hashing (CPH) to solve concept drift and reduce computational cost in image retrieval.
• To maintain consistent hash codes for evolving image concepts over time.

Main Methods:

• CPH learns new hash functions that generate the same hash codes for both old and new images of a concept.
• The CPH objective function includes isomorphic similarity, hash code partition balancing, and heterogeneous similarity fitness.

Main Results:

• CPH demonstrated superior retrieval precisions compared to existing methods across 11 concept drift scenarios.
• CPH effectively preserves original hash codes, eliminating the need to update previously stored images.

Conclusions:

• CPH offers an efficient and effective solution for image retrieval in dynamic databases with concept drift.
• This method significantly enhances retrieval accuracy and reduces computational overhead in evolving datasets.