Supervised Adaptive Similarity Matrix Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Compact hash codes are crucial for large-scale multimedia retrieval, reducing storage and computation demands.
• Existing hashing methods often rely on predefined similarity matrices, which may not accurately represent true image relationships, leading to suboptimal performance, especially in multi-label and zero-shot scenarios.

Purpose of the Study:

• To propose a novel supervised hashing method, Supervised Adaptive Similarity Matrix Hashing (SASH), to address the limitations of predefined similarity matrices.
• To enhance the accuracy of similarity relationships in hashing for improved multimedia retrieval.

Main Methods:

• SASH adaptively learns the similarity matrix by maintaining consistency between the feature and label spaces.
• It extracts label correlations to optimize the similarity matrix, improving the representation of image similarities.

Main Results:

• Experiments on multiple large-scale benchmark datasets, including multi-label and zero-shot datasets, demonstrate SASH's effectiveness.
• SASH significantly outperforms several state-of-the-art hashing techniques in retrieval performance.

Conclusions:

• SASH offers a superior approach to supervised hashing by adaptively learning similarity matrices through feature-label space consistency.
• The method shows excellent performance, particularly for datasets where accurate similarity representation is critical.