Cross-View Retrieval via Probability-Based Semantics-Preserving Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Hashing methods accelerate nearest neighbor search in large-scale data.
• Cross-view retrieval from multiview data presents unique challenges for efficient querying.

Purpose of the Study:

• To propose an effective probability-based semantics-preserving hashing (SePH) method for cross-view retrieval.
• To generate unified hash codes for instances across multiple views, preserving semantic consistency.

Main Methods:

• SePH transforms semantic affinities into a probability distribution for training.
• Minimizes Kullback-Leibler divergence between probability distributions in Hamming space.
• Employs predictive models for learning hash functions and a probabilistic approach for out-of-sample extension.

Main Results:

• SePH generates a single hash code for all views of an instance.
• Learned hash functions project view-specific features into hash codes.
• Experimental results on benchmark datasets confirm SePH's effectiveness and reasonableness.

Conclusions:

• The proposed SePH method offers an effective solution for cross-view retrieval in large-scale multiview data.
• SePH's probabilistic approach ensures semantic consistency and improves retrieval efficiency.