One key challenge in zero-shot classification（ZSC） is the exploration of knowledge hidden in unseen classes. Generative methods such as generative adversarial networks（GANs） are typically employed to generate the visual information of unseen classes. However, the majority of these methods exploit global semantic features while neglecting the discriminative differences of local semantic features when synthesizing images, which may lead to sub-optimal results. In fact, local semantic information can provide more discriminative knowledge than global information can. To this end, this paper presents a new triple discriminator GAN for ZSC called TDGAN, which incorporates a text-reconstruction network into a dual discriminator GAN（D2GAN）, allowing to realize cross-modal mapping from text descriptions to their visual representations. The text-reconstruction network focuses on key text descriptions for aligning semantic relationships to enable synthetic visual features to effectively represent images. Sharma-Mittal entropy is exploited in the loss function to make the distribution of synthetic classes be as close as possible to the distribution of real classes. The results of extensive experiments over the Caltech-UCSD Birds-2011 and North America Birds datasets demonstrate that the proposed TDGAN method consistently yields competitive performance compared to several state-of-the-art ZSC methods.
展开▼
