Deep Softtriple hashing for Multi-Label cross-modal retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Information Retrieval

Background:

• Deep cross-modal hashing methods often use mini-batch training, limiting neighborhood exploration and embedding quality, especially on multi-label data.
• Existing SoftMax loss optimizations model classes with single centers, failing to capture intra-class semantic clusters and increasing the semantic gap.

Purpose of the Study:

• To propose Deep SoftTriple Hashing (DSTH), a novel framework for learning compact hash codes that preserve semantic similarities in cross-modal retrieval.
• To address limitations in current deep hashing methods by modeling intra-class variations and reducing semantic gaps.

Main Methods:

• Introduced a multi-center strategy for each class to model heterogeneous sample distributions and reduce intra-class variance.
• Developed a class-center strategy with L2,1 regularization to aggregate similar centers and determine an optimal number of centers.
• Incorporated a semantic position quantization loss to minimize quantization error and improve binary code discriminability.

Main Results:

• DSTH achieved significant improvements in cross-modal retrieval, with absolute mAP gains of 1.2%-9.3% over strong baselines on three multi-label datasets.
• Demonstrated superior performance on Precision-Recall curves compared to existing methods.
• Effectively learned compact hash codes that better preserve semantic similarities.

Conclusions:

• DSTH effectively models intra-class local structures using multiple centers, outperforming single-center approaches in complex multi-label retrieval scenarios.
• The proposed framework enhances embedding learning and reduces the semantic gap, leading to state-of-the-art performance in cross-modal retrieval.
• DSTH offers a promising approach for improving large-scale cross-modal retrieval systems.