Incremental Hash-Bit Learning for Semantic Image Retrieval in Nonstationary Environments.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semantic image retrieval faces challenges due to concept drift and distribution changes in image data over time.
• Static hashing methods trained on fixed datasets are inadequate for non-stationary semantic image retrieval.
• Retraining entire hash tables for new data is computationally inefficient.

Purpose of the Study:

• To propose a novel incremental hash-bit learning method for semantic image retrieval in non-stationary environments.
• To develop an efficient approach for updating image retrieval models with evolving data.
• To improve the accuracy and relevance of semantic image retrieval results.

Main Methods:

• Introduced an incremental hash-bit learning method that iteratively selects and trains new hash bits.
• Utilized a 3-component objective function to evaluate hash bits based on information preservation, partition balancing, and bit angular difference.
• Combined knowledge from existing and newly trained hash bits to adapt to concept drift.
• Developed a re-ranking mechanism using weighted hash bits for improved retrieval.

Main Results:

• The proposed method demonstrated superior performance compared to stationary, table-based incremental, and online hashing methods.
• Experimental results were validated across 15 different simulated non-stationary data environments.
• The method automatically adjusted the number of old and new data bits based on concept drift.

Conclusions:

• The incremental hash-bit learning method effectively handles concept drift in semantic image retrieval.
• This approach offers an efficient and adaptive solution for updating image retrieval systems.
• The proposed method significantly enhances semantic image retrieval performance in dynamic data scenarios.