Zero-Shot Learning via Category-Specific Visual-Semantic Mapping and Label Refinement.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-Shot Learning (ZSL) classifies unseen categories using seen categories via visual-semantic mapping.
• A key challenge in ZSL is the projection domain shift, where mappings learned on seen categories fail to generalize to unseen ones.

Purpose of the Study:

• To address the projection domain shift problem in Zero-Shot Learning.
• To develop methods that learn adaptive visual-semantic mappings for improved generalization to unseen categories.

Main Methods:

• Proposed Adaptive Embedding ZSL (AEZSL) for category-specific adaptive mapping and progressive label refinement.
• Introduced Deep AEZSL (DAEZSL), a deep adaptive embedding model trained once for arbitrary unseen categories, addressing large-scale ZSL.

Main Results:

• AEZSL and DAEZSL effectively mitigate the projection domain shift issue.
• Both methods achieved state-of-the-art performance on image classification tasks.
• Evaluated on three small-scale and one large-scale benchmark datasets.

Conclusions:

• The proposed adaptive embedding approaches significantly enhance Zero-Shot Learning performance.
• DAEZSL offers an efficient solution for large-scale ZSL by enabling single-time training for broad applicability.
• Category-specific visual-semantic mapping is crucial for robust ZSL generalization.