Learning Autoencoder of Attribute Constraint for Zero-Shot Classification

The goal of zero-shot classification (ZSC) is to classify target classes precisely based on learning a semantic mapping from a feature space to a semantic knowledge space. However, the learned semantic mapping is only concerned with predicting source classes. Applying the semantic mapping to target classes directly will suffer from the semantic shift problem. In this paper, we propose a novel method called autoencoder of attribute constraint (AOAC) to settle this problem. In AOAC, we adopt the encoder-decoder paradigm to learn the semantic mapping. Additionally, we take the inaccurate attributes of source images into consideration and generate virtual data to solve it. The experimental results on two challenging datasets show that our proposed AOAC can resolve the semantic shift problem effectively and also improve the computational speed significantly.