Zero-shot learning via categorization-relevant disentanglement and discriminative samples synthesis

Zero-shot learning (ZSL) trains classifiers on seen classes and utilizes class-level semantic attributes to recognize unseen classes. In recent years, feature generation methods have been employed to address the data imbalance issue in ZSL by synthesizing samples of unseen classes. Nevertheless, the current generation methods solely rely on training the generator using seen classes, resulting in synthesized visual features that are inevitably biased toward seen classes and lack discriminative information. In this paper, we propose an effective ZSL method, termed zero-shot learning via categorization-relevant disentanglement and discriminative samples synthesis. To mitigate the category bias problem, we incorporate a domain discriminator and a semantic decoder into the generative network. By maximizing the decision boundary between categories, distinguishing unseen classes and promoting the alignment of synthetic features with their corresponding semantic embeddings, to enhance semantic consistency. In addition, to reduce the interference of redundant information in visual features, we propose a batch recombining-based disentangling method. The original visual features are projected into categorization-irrelevant and categorization-relevant representations through a conditional encoder, and a correlation penalty is employed to ensure the independence between the two component representations. The categorization-irrelevant representation is reorganized and spliced with the categorization-relevant representation, and disentanglement is achieved by calculating the difference between the post-reorganized and pre-reorganized representations. Finally, we train a classifier using the categorization-relevant representations to improve classification accuracy. Experiments are conducted on four publicly available ZSL datasets, and the proposed method achieves superior results, demonstrating its effectiveness in addressing the challenges of ZSL.