Attribute-Guided Network for Cross-Modal Zero-Shot Hashing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot hashing (ZSH) enables models trained on seen categories to generalize to unseen ones, typically using semantic embeddings.
• Current ZSH research predominantly focuses on single-modal retrieval, like image-based image retrieval (IBIR).
• Cross-modal retrieval, however, is more prevalent in real-world applications, presenting a significant research gap.

Purpose of the Study:

• To develop a novel network for cross-modal zero-shot hashing (CMZSH) that addresses limitations of existing single-modal approaches.
• To enable retrieval across different modalities, specifically image-based image retrieval (IBIR) and text-based image retrieval (TBIR).
• To bridge the modality gap and zero-shot setting by aligning diverse data into a unified semantic attribute space.

Main Methods:

• Propose the attribute-guided network (AgNet) designed for CMZSH.
• Align image and text data into a shared, semantically rich attribute space to handle modality heterogeneity.
• Develop a strategy to leverage attributes for guiding hash code generation for both image and text within a single network.

Main Results:

• AgNet demonstrates superior performance on both cross-modal and single-modal zero-shot image retrieval tasks.
• Experiments conducted on benchmark datasets (AwA, SUN, ImageNet) validate the effectiveness of the proposed method.
• The attribute-guided approach successfully bridges the gap between different data modalities in a zero-shot context.

Conclusions:

• The proposed AgNet effectively addresses the CMZSH task, outperforming existing methods.
• Aligning data into an attribute space is a viable strategy for handling modality heterogeneity in zero-shot learning.
• AgNet offers a unified framework for both single-modal and cross-modal zero-shot retrieval.