Semantic-Guided Prompt Learning Network for Generalized Zero-Shot Learning

Generalized zero-shot learning (GZSL) addresses the challenge of recognizing both seen and unseen classes with only training data from the seen classes. While the large-scale model CLIP holds promise for GZSL, a significant obstacle remains: the scarcity of high-quality prompts. To overcome this, we present a novel prompt learning approach for GZSL that leverages semantic information to guide the construction of a generic prompt template applicable to both seen and unseen classes. Specifically, we propose a semantic-guided prompt tuning network, which effectively learns the prompt template using semantic knowledge, enabling its utilization across seen and unseen classes. We extensively evaluate our approach on three GZSL datasets, where our network consistently achieves competitive performance across all three datasets. By addressing the challenge of prompt quality, our method demonstrates the potential of CLIP in GZSL tasks and highlights the importance of semantic guidance in learning effective prompt templates.