Zero-shot Learning via the fusion of generation and embedding for image recognition

Abstract

Traditional image recognition methods require abundant labelled images to learn robust recognition models, and only the images from the classes that appear in the training set (also called seen classes) can be recognized. Thus, determining how to recognize an image from a novel class (also called unseen class) is very challenging. Zero-Shot Learning (ZSL) is proposed to address the above issue. In this paper, we present a novel ZSL model called ZSL based on class prototype and dual latent subspace learning with reconstruction (ZSL-CPLSR). Aiming at the problems of domain shift and information loss, ZSL-CPLSR integrates generation and embedding into a unified framework in the inductive setting, which takes full advantage of the semantic information of unseen classes to alleviate the domain shift and learns more discriminative information with less effective information loss. We conduct comprehensive experiments on four benchmark datasets widely adopted in ZSL. The experimental results show the superiority of ZSL-CPLSR to the state-of-the-art ZSL methods, validating the effectiveness of ZSL-CPLSR.