Generalized zero-shot learning based on dual latent space reconstruction

Abstract

Generalized Zero-Shot Learning (GZSL) is characterized as a training process that comprises visual samples from seen classes and semantic samples from seen and unseen classes, followed by a testing process that classifies visual samples from seen and unseen classes. Existing zero-shot learning (ZSL) approaches suffer from domain shift and information loss issues as a result of class differences between visible and unseen classes, as well as uneven image distribution. In this study, a generalized zero-shot learning strategy based on dual latent space reconstruction (DLR-GZSL) is proposed. The method aims to establish a latent space of shared semantic and visual information, uses dual learning to align different modal representations to alleviate the domain shift problem, uses triplet loss to improve intra-class diversity and inter-class separability of the generated samples, and uses information bottleneck to retain as much valid generated feature information as possible to reduce information loss. Experiments on the CUB, SUN, AWA1, and AWA2 datasets reveal that the suggested method has more accurate than previous methods, demonstrating its effectiveness.