Random Online Hashing for Cross-Modal Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Supervised cross-modal hashing methods are crucial for efficient searching in large multimedia databases.
• Existing methods using similarity matrices or matrix factorization face limitations in scalability and preserving modality-specific information.
• There is a need for advanced methods that can effectively integrate semantic correlations and modality-specific features.

Purpose of the Study:

• To propose a novel supervised cross-modal hashing method, Random Online Hashing (ROH).
• To address the limitations of existing methods in preserving semantic information and modality-specific features.
• To develop an efficient method for handling real-time data dynamically.

Main Methods:

• Introduced a linear bridging strategy to simplify pair-wise similarities factorization into linear optimization.
• Utilized a bridging matrix to establish a bidirectional linear relation between hash codes and labels.
• Proposed a Maximum Eigenvalue Direction (MED) embedding method to preserve critical modality-specific information and adopted an online structure for dynamic data handling.

Main Results:

• The bridging matrix effectively preserves semantic similarities and reduces semantic distances between similar labeled samples.
• The MED embedding method successfully embeds critical information into modality-specific hash codes.
• Extensive experiments on benchmark datasets show ROH outperforms existing state-of-the-art cross-modal hashing methods.

Conclusions:

• ROH offers an effective solution for supervised cross-modal hashing by improving semantic information preservation and modality-specific feature integration.
• The proposed linear bridging strategy and MED embedding method enhance the performance and scalability of cross-modal hashing.
• The online structure makes ROH suitable for real-time applications with dynamic data streams.