Learning Autoencoder of Attribute Constraint for Zero-Shot Classification

The goal of zero-shot classification (ZSC) isto classify target classes precisely based on learning asemantic mapping from a feature space to a semanticknowledge space. However, the learned semantic mappingis only concerned with predicting source classes. Applyingthe semantic mapping to target classes directly will sufferfrom the semantic shift problem. In this paper, we proposea novel method called autoencoder of attribute constraint(AOAC) to settle this problem. In AOAC, we adopt theencoder-decoder paradigm to learn the semantic mapping.Additionally, we take the inaccurate attributes of sourceimages into consideration and generate virtual data to solveit. The experimental results on two challenging datasetsshow that our proposed AOAC can resolve the semanticshift problem effectively and also improve the computationalspeed significantly.