Deep Variational and Structural Hashing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional deep hashing methods often rely on standard convolutional and fully-connected layers.
• Existing methods may not optimally capture latent feature representations for multimedia retrieval.
• Scalable cross-modal retrieval presents challenges in learning unified representations.

Purpose of the Study:

• To propose a novel deep hashing method (DVStH) for learning compact binary codes.
• To develop a probabilistic framework for inferring latent feature representations.
• To extend the method for scalable cross-modal multimedia retrieval (CM-DVStH).

Main Methods:

• Developed a deep variational and structural hashing (DVStH) method using a probabilistic framework.
• Introduced a 'struct layer' for binary code generation instead of a bottleneck hash layer.
• Extended DVStH to CM-DVStH with a deep fusion network for image-text correlation and modality-specific networks for out-of-sample extension.

Main Results:

• The proposed DVStH method effectively learns discriminative and generative binary codes.
• CM-DVStH demonstrates efficacy in cross-modal scalable multimedia retrieval.
• Experimental results on five benchmark datasets validate the proposed approach's performance.

Conclusions:

• The DVStH method offers an effective approach for compact binary code learning in multimedia retrieval.
• CM-DVStH successfully addresses cross-modal retrieval challenges by maximizing image-text correlations.
• The proposed hashing techniques show significant potential for enhancing multimedia retrieval systems.