Fast Partial-Modal Online Cross-Modal Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Cross-Modal Hashing (CMH) is vital for large-scale retrieval but struggles with real-time adaptation to streaming data.
• Existing online CMH methods face challenges with partial-modal data and high retraining costs.

Purpose of the Study:

• To develop an efficient online CMH method for streaming partial-modal data.
• To address the limitations of existing CMH models in dynamic, incomplete data environments.

Main Methods:

• Proposed Fast Partial-modal Online Cross-Modal Hashing (FPO-CMH).
• Introduced a multimodal dual-tier anchor bank for seamless adaptation to partial data.
• Utilized gradient accumulation, asynchronous optimization, and initial-anchor rehearsal for efficient online learning and preventing catastrophic forgetting.

Main Results:

• FPO-CMH demonstrates superior performance in handling streaming partial-modal multimodal data.
• The method effectively adapts pre-trained CMH models to new, incomplete data streams.
• Achieved efficient online learning without frequent hash function retraining or database hash code updates.

Conclusions:

• FPO-CMH offers a robust solution for real-time cross-modal retrieval with streaming partial-modal data.
• The approach provides a significant advancement over existing CMH techniques in realistic scenarios.
• Enables efficient and cost-effective adaptation of CMH models to evolving data streams.