Content-Attribute Disentanglement for Generalized Zero-Shot Learning

Humans can recognize or infer unseen classes of objects using descriptions explaining the characteristics (semantic information) classes. However, conventional deep learning models trained in a supervised manner cannot classify that were during training. Hence, many studies have been conducted into generalized zero-shot (GZSL), which aims to produce system both seen and classes, by transferring learned knowledge from Since share common semantic space, extracting appropriate information images is essential for GZSL. In addition semantic-related (attributes), also contain semantic-unrelated (contents), degrade classification performance model. Therefore, we propose content-attribute disentanglement architecture separates content attribute images. The proposed method comprised three major components: 1) feature generation module synthesizing visual features; 2) discriminating codes images; 3) an comparator measuring compatibility between class prototypes act as ground truth.With extensive experiments, show our achieves state-of-the-art competitive results on four benchmark datasets Our outperforms existing methods all datasets. Moreover, has best accuracy well retrieval task.