Cross Knowledge-based Generative Zero-Shot Learning approach with Taxonomy Regularization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) enables recognition of unseen classes but faces cross-modality and cross-domain challenges.
• Existing ZSL methods struggle with generalizing to new data distributions.

Purpose of the Study:

• To develop a novel generative network-based ZSL approach to address cross-modality and cross-domain limitations.
• To improve the performance of ZSL in image classification and retrieval tasks.

Main Methods:

• A generative network is employed to synthesize visual features from semantic features.
• The Cross Knowledge Learning (CKL) scheme is proposed for better semantic-to-visual feature embedding.
• Taxonomy Regularization (TR) is introduced to generate more generalized visual features for improved unseen image recognition.

Main Results:

• The proposed approach significantly improves the ability to recognize unseen classes.
• Experiments on benchmark datasets (AwA1, AwA2, CUB, NAB, aPY) demonstrate superior performance over state-of-the-art methods.
• The method shows enhanced capabilities in both ZSL image classification and retrieval.

Conclusions:

• The developed generative network-based ZSL approach effectively alleviates cross-modality and cross-domain challenges.
• The integration of CKL and TR leads to more robust and generalized visual features for ZSL.
• This work offers a promising direction for advancing zero-shot learning capabilities.