Attribute alignment networks for generalized zero-shot learning

Part-based embedding methods with attention mechanism achieved outstanding results in zero-shot learning (ZSL). However, affected by intra-class variations in the datasets (i.e., different samples of the same class present different attribute characteristics), it is difficult for models based on traditional attention mechanisms to achieve accurate visual-attribute alignment. To tackle this problem, we propose a novel approach to fully utilize attributes information, referred to as attribute alignment networks (AAN). It consists of the attribute alignment (AA) pipeline and the attribute enhancement (AE) module. AA pipeline is a brand-new solution for part-based embedding method, which realizes visual-attribute alignment in both attribute space and attribute semantic space under the supervision of class attribute vectors and attribute word vectors, respectively. AE module employs the Graph Neural Networks (GNNs) to project visual features to the attribute semantic space. Based on the constructed attribute relation graph (ARG) and self-attention mechanism, AE module generates the enhanced representation of attributes to minimize the influence of intra-class variations. Experiments on standard datasets demonstrate that the enhanced attribute representation greatly improves the classification performance. Overall, AAN outperforms the other state-of-the-art performances in ZSL and GZSL tasks.