Zero-shot learning (ZSL) aims to learn a classifier for unseen classes by exploiting both training data from seen classes and external knowledge. In many visual tasks such as image classification, a set of high-level attributes that describe the semantic properties of classes are used as the external knowledge to bridge seen and unseen classes. While the attributes are usually treated equally by previous ZSL studies, we observe that the contribution of different attributes varies significantly over model training. To adaptively exploit the discriminative information embedded in different attributes, we propose a novel encoder–decoder framework with attention mechanism on the attribute level for zero-shot learning. Specifically, by mapping the visual features into a semantic space, the more discriminative attributes are emphasized with larger attention weights. Further, the attentive attributes and the class prototypes are simultaneously decoded to the visual space so that the hubness problem can be eased. Finally, the labels are predicted in the visual space. Extensive experiments on multiple benchmark datasets demonstrate that our proposed model achieves a significant boost over several state-of-the-art methods for ZSL task and comparative results for generalized ZSL task.

零样本学习 (ZSL) 旨在通过利用来自已见类的训练数据和外部知识来学习未见类的分类器。在图像分类等许多视觉任务中，一组描述类语义属性的高级属性被用作外部知识，以连接可见类和不可见类。虽然之前的 ZSL 研究通常对这些属性进行同等对待，但我们观察到不同属性的贡献在模型训练中差异很大。为了自适应地利用嵌入在不同属性中的判别信息，我们提出了一种新颖的编码器 - 解码器框架，在属性级别具有注意机制，用于零样本学习。具体来说，通过将视觉特征映射到语义空间，用更大的注意力权重强调更具辨别力的属性。此外，注意力属性和类原型同时解码到视觉空间，从而可以缓解中心问题。最后，在视觉空间中预测标签。在多个基准数据集上进行的大量实验表明，我们提出的模型在 ZSL 任务的几种最先进方法和广义 ZSL 任务的比较结果上取得了显着的提升。