Learning Internal Semantics with Expanded Categories for Generative Zero-Shot Learning

In recent years, generative Zero-Shot Learning (ZSL) has attracted much attention due to its better performance than traditional embedding methods. Most generative ZSL methods exploit category semantic plus Gaussian noise to generate visual features. However, there is a contradiction between the unity of category semantic and the diversity of visual features. The semantic of a single category cannot accurately correspond to different individuals in the same category. This is due to the different visual expression of the same category. Therefore, to solve the above mentioned problem we propose a novel semantic augmentation method, which expands a single semantic to multiple internal sub-semantics by learning expanded categories, so that the generated visual features are more in line with the real visual feature distribution. At the same time, according to the theory of Convergent Evolution, the sub-semantics of unseen classes are obtained on the basis of the expanded semantic of their similar seen classes. Four benchmark datasets are employed to verify the effectiveness of the proposed method. In addition, the category expansion is also applied to three generative methods, and the results demonstrate that category expansion can improve the performance of other generative methods. Code is available at: https://github.com/njzxj/EC-GZSL.