Three-Stage Semisupervised Cross-Modal Hashing With Pairwise Relations Exploitation.

Area of Science:

• Computer Science
• Information Retrieval

Background:

• Hashing methods offer efficient storage and computation for cross-modal retrieval.
• Supervised hashing excels with labeled data but annotation is costly.
• Existing semisupervised methods often learn pseudolabels, hash codes, and functions simultaneously, posing optimization challenges.

Purpose of the Study:

• To propose a novel semisupervised hashing method (TS3H) that efficiently handles both labeled and unlabeled data for cross-modal retrieval.
• To address the limitations of expensive data annotation in supervised hashing methods.
• To develop a cost-effective and precise optimization approach for semisupervised hashing.

Main Methods:

• A three-stage approach: 1. Classifier learning using supervised information to predict pseudolabels for unlabeled data. 2. Hash code learning by unifying provided and predicted labels, leveraging pairwise relations. 3. Modality-specific hash function generation.
• Utilizes pairwise relations to supervise both classifier and hash code learning stages.
• Decomposes the learning process into distinct, individually optimized stages for precision and cost-effectiveness.

Main Results:

• The proposed TS3H method demonstrates superior efficiency and performance compared to state-of-the-art shallow and deep cross-modal hashing methods.
• Experimental results on benchmark databases validate the effectiveness of the TS3H approach.
• The method successfully integrates labeled and unlabeled data for improved cross-modal retrieval.

Conclusions:

• TS3H offers a viable and effective solution for cross-modal retrieval by overcoming the data annotation bottleneck.
• The staged optimization strategy ensures precise and cost-effective learning of hash codes and functions.
• The method shows significant potential for real-world applications requiring efficient cross-modal data handling.