Learning discriminative and representative feature with cascade GAN for generalized zero-shot learning

Abstract

Zero-Shot Learning (ZSL) aims to employ seen images and their related semantics to identify unseen images through knowledge transfer. Among past numerous methods, the generative methods are more prominent and achieve better results than other methods. However, we find the input for generating samples is too monotonous, there are only semantics of each class and artificially defined noise, which makes the generated visual features non-discriminative and the classifier cannot effectively distinguish them. In order to solve this problem, we propose a novel approach with cascade Generative Adversarial Network (GAN) to generate discriminative and representative features. In this method, we define a latent space where the features from different categories are orthogonal to each other and the generator for this latent space is learned with a Wasserstein GAN. In addition, in order to make up for the deficiency that the features in this latent space cannot accurately simulate the true distribution of species, we utilize another Wasserstein GAN or Cramér GAN cascaded with the previous one to generate more discriminative and representative visual features. In this way, we can not only expand the content used as input in the generation process, but also make the final generated visual features clear and separable under the influence of latent spatial orthogonality. Extensive experiments on five benchmark datasets, i.e., AWA1, AWA2, CUB, SUN and APY, demonstrate that our proposed method can outperform most of the state-of-the-art methods on both conventional and generalized zero-shot learning settings.