Augmented semantic feature based generative network for generalized zero-shot learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) enables object recognition without direct training data.
• Generalized zero-shot learning (GZSL) handles both seen and unseen classes during testing.
• Existing generative methods for ZSL often produce visually limited features due to incomplete semantic information.

Purpose of the Study:

• To enhance semantic features by incorporating discriminative visual information.
• To develop a novel generative network, ASFGN, for synthesizing separable visual representations of unseen classes.
• To improve the training stability and generalization of generative models in ZSL.

Main Methods:

• Utilizing discriminative visual features to augment user-defined semantic information.
• Proposing the Augmented Semantic Feature Based Generative Network (ASFGN) for synthesizing visual features.
• Introducing a novel collapse-alleviate loss function to stabilize GAN training and prevent mode collapse.

Main Results:

• The proposed method significantly improves the quality and separability of synthesized visual features.
• ASFGN demonstrates state-of-the-art performance on benchmark datasets for both ZSL and GZSL.
• The collapse-alleviate loss enhances training stability and the generalization capabilities of the generative network.

Conclusions:

• Augmenting semantic features with visual information is crucial for effective ZSL.
• ASFGN provides a robust framework for generating diverse and separable visual features for unseen classes.
• The proposed approach advances the capabilities of zero-shot learning in computer vision.