Low-rank Multi-modal Features Fusion Semantic Autoencoder For Zero-shot Learning

In the context of the continuous increase of new things, zero-shot learning (ZSL) has been proposed to reduce recognition costs, with the goal of classifying or predicting new classes that were not seen during training. Previous research introduced the Semantic Autoencoder (SAE), however, this embedding-based approach is prone to domain shift issues. To address this problem, this paper proposes a Low-Rank Multi-modal Feature Fusion Semantic Autoencoder (LMFFSAE), where the encoder projects multi-modal features into a latent representation space, and the decoder reconstructs visual semantic features. The learned projection function has less domain specificity. Additionally, the model introduces a low-rank constraint to highlight shared attributes between different classes, thereby mitigating the domain shift issue. Experiments on four zero-shot learning benchmark datasets demonstrate that LMFFSAE outperforms existing methods, proving it to be an effective zero-shot learning approach.