Zero-Shot Learning Based on Quality-Verifying Adversarial Network

Abstract

Recently, generative adversarial network (GAN)-based zero-shot learning methods have attracted widespread attention. However, due to the randomness of GAN generation, most existing methods cannot well guarantee to generate sufficiently reliable features and have good generalization ability. Targeting at these problems, we propose an effective Quality-Verifying Adversarial Network (QVAN) that consists of one generator and double discriminators. Adversarial learning between the former discriminator and generator is to generate visual features, which can be partitioned into pseudo-generated features and reliable-generated features. The latter discriminator is used for quality-verifying that will guide the generator to generate more reliable features that are near the real visual features. To avoid over-fitting and ensure intra-class diversity, we set the threshold for each class to distinguish pseudo-generated features and reliable-generated features. To further preserve both compactness and discriminability of the samples, we introduce the class metric constraint, which are more conducive to classification. Moreover, we introduce <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\ell _{1,2}$</tex-math></inline-formula> -norm constraint to fully consider the specific distribution among different classes, thus making the generated features more discriminant. Extensive experiments on several real-world datasets show the effectiveness of the proposed approach, which demonstrate the advantage over the state-of-the-art methods.