关键词: 零样本学习, 语义拓展, 视觉-语义映射, 对比学习

Abstract:  In zero-shot image classification， semantic embedding technology （i.e.， using semantic attributes to describe class labels） provides the means to generate visual features for unknown objects by transferring knowledge from known objects. Current research often utilizes class semantic attributes as auxiliary information for describing class visual features. However， class semantic attributes are typically obtained through external paradigms such as manual annotation， resulting in weak consistency with visual features. Moreover， a single class semantic attribute is insufficient to capture the diversity of visual features. To enhance the diversity of class semantic attributes and their capacity to describe visual features， this paper introduces a Semantic Extension and Embedding for Zero-Shot Learning （SeeZSL） based on semantic extension and embedding. SeeZSL expands semantic information by constructing a latent semantic space for each class， enabling the generation of visual features for unknown classes using this semantic space. Additionally， to address the issues of weak consistency and the lack of discriminative ability between the original feature space and class semantic attributes， a semantic extension-based generation model is integrated with an contrastive-embedding model. The effectiveness of the proposed SeeZSL method was experimentally validated on four benchmark datasets.

Key words: zero-shot learning, semantic expansion, visual-semantic mapping, contrastive learning