Dual insurance for generalized zero-shot learning

Traditional zero-shot learning aims to use the trained model to accurately classify samples from unseen classes, while for the more difficult task of generalized zero-shot learning, the trained model needs to classify samples from both seen and unseen classes into the correct classes. Because only seen class samples are available during training, generalized zero-shot learning meets great challenges in classification. Generative model is one of the good methods to solve this problem. However, the samples generated by the generative model are often of poor quality. In addition, there are semantic redundancies in the generated samples that are not conducive to classification. To solve these problems, we proposed the dual insurance model (DI-GAN) for generalized zero-shot learning in this paper, including a feature generation module and a semantic separation module. They guarantee the high quality of generated features and the good classification performance respectively. Specifically, the first insurance is based on generative adversarial network, whose generator is constrained by a clustering method to make the generated samples close to the real samples. The second insurance is based on variational autoencoder, including semantic separation, instance network and classification network. Semantic separation is designed to extract the semantically related parts which are beneficial to classification, while instance network acting on the semantically related parts is used to ensure the classification performance. Extensive experiments on four benchmark datasets show the competitiveness of the proposed DI-GAN.