FAVC: Fine-Grained Attribute-Aware Visual Cross-Modal for Zero-Shot Learning

Zero-shot learning (ZSL) aims to predict the categories of unseen classes by transferring seen class knowledge to unseen classes. Attribute text descriptions as annotations of visual features are currently one of the most effective and widely used semantic information in ZSL. However, conventional vision-language pre-trained models only focus on the global featurealignment between images and text, while ignoring the more fine-grained semantic alignment between local visual features and text phrases, which restricts the model’s ability to zero-shot learning. Therefore, we propose a Fine-Grained Attribute-Aware Visual Cross-Modal (FAVC) learning framework, which correlates local visual features with attribute text descriptions via a cross-attention model and enhances the feature extraction capability and zero-shot recognition ability of the model through a contrastive learning training paradigm. Three standard ZSL benchmark tests show that FAVC achieves excellent performanceand has a better ability to capture localized regions with significant features in images.