Incremental Hashing for Semantic Image Retrieval in Nonstationary Environments

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• The internet hosts a vast and ever-growing volume of images daily.
• Existing image retrieval hashing methods are primarily designed for static databases and struggle with evolving data distributions.
• Concept drift and the emergence of new image classes degrade retrieval performance in dynamic datasets over time.

Purpose of the Study:

• To address the limitations of static hashing methods in dynamic image databases.
• To propose a novel incremental hashing (ICH) method capable of handling evolving data distributions.
• To improve the long-term retrieval performance of large-scale image databases.

Main Methods:

• The proposed incremental hashing (ICH) method utilizes multihashing to incorporate knowledge from newly arriving images.
• A weight-based ranking mechanism is employed to ensure retrieval results adapt to the current data environment.
• ICH is designed to efficiently manage changes in data distribution without full database retraining.

Main Results:

• Experimental results demonstrate the effectiveness of the ICH method in managing dynamic changes within image databases.
• The ICH approach mitigates performance degradation caused by concept drift and the addition of new image classes.
• ICH shows improved adaptability and retrieval accuracy compared to traditional methods in evolving datasets.

Conclusions:

• The incremental hashing (ICH) method offers a robust solution for image retrieval in dynamic online environments.
• ICH successfully handles both new data additions and concept drift, enhancing retrieval system adaptability.
• This approach provides a computationally efficient and effective way to maintain high retrieval performance as databases evolve.