Generalized zero-shot learning (GZSL) models, trained on images from seen classes, are targeted to classify new images from both the seen and unseen classes. Previous works take the original semantic description as the input without further exploring its intrinsic information, which may hinder the knowledge transfer from the seen classes to the unseen classes. To facilitate knowledge transfer, a new approach is proposed in this paper to view the GZSL as a multilabel classification task. The labels, obtained from semantic attributes, are the similarities between seen and unseen classes. The proposed attribute attention module in our network is capable of extracting and enhancing the discriminative parts of image features. Moreover, two loss functions with pseudo data generation are implemented in the network to balance the transferability and the discriminability. Comprehensive evaluations show that the proposed approach outperforms the stateof-the-art methods in three datasets, APY, AWA2 and SUN, in the GZSL setting.
分享
引用
