Learning exclusive discriminative semantic information for zero-shot learning

Jian-Xun Mi, Zhonghao Zhang, Debao Tai, Li-Fang Zhou, Wei Jia

Abstract: Zero-shot learning (ZSL) aims to recognize unseen classes relying on the knowledge transferred from seen categories. This study presents new methods to solve two main challenges in ZSL. First, as human-annotated semantics are not discriminative enough to identify unseen classes, we propose constructing a novel latent semantic space based on the semantic attributes and designing a class-wise classifier with class-specific information maximize the discrimination of the latent semantics. Besides, to alleviate the common space’s semantic overlapping problem, we first propose constructing exclusive latent class prototypes by exclusive lasso (EL). Second, since previous ZSL methods learn visual-semantic projection between visual features and corresponding single class-level semantics directly, i.e., one-vs-all projection, which neglects the interference caused by background and noises in the image, we leverage the simple quadratic regression to soften this hard constraint. The proposed new model also alleviates the inherent domain shift problem by adopting the dual semantic auto-encoder to connect visual space, semantic space, and latent space, respectively. Comprehensive experiments on five benchmark datasets demonstrate the effectiveness of the proposed model.