Bidirectional Semantic Consistency Guided Contrastive Embedding for Generative Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generative zero-shot learning (GZSL) methods synthesize features for unseen classes using image and semantic information.
• Existing GZSL methods often fail to ensure semantic consistency in synthesized features due to reliance on global image features, leading to poor discriminative power.
• Bias in knowledge transfer from seen to unseen classes remains a challenge in GZSL.

Purpose of the Study:

• To propose a novel Bidirectional Semantic Consistency Guided (BSCG) generation model for improved GZSL.
• To enhance the semantic consistency and discriminative power of synthesized features for unseen classes.
• To ensure robust knowledge transfer and improve the generalization ability of GZSL models.

Main Methods:

• The proposed BSCG model employs a Bidirectional Semantic Guidance Framework (BSGF).
• BSGF integrates Attribute-to-Visual Guidance (AVG) and Visual-to-Attribute Guidance (VAG) for enhanced visual-semantic interaction.
• A Contrastive Consistency Space (CCS) is introduced to optimize feature quality by increasing intra-class compactness and inter-class separability.

Main Results:

• The BSCG model demonstrates significant performance improvements over state-of-the-art approaches.
• Experiments were conducted on three benchmark datasets for both conventional and generalized zero-shot learning settings.
• The proposed methods effectively address limitations in semantic consistency and feature discriminability.

Conclusions:

• The BSCG model offers a robust solution for generative zero-shot learning by ensuring semantic consistency.
• The integration of BSGF and CCS leads to superior knowledge transfer and generalization capabilities.
• The approach shows strong potential for advancing research in zero-shot learning.