2025 Volume 29 Issue 5 Pages 1029-1038
Generalized zero-shot learning (GZSL) focuses on recognizing classes, both seen and unseen, without the need for labeled data specifically for the unseen classes. Generative GZSL has attracted considerable attention because it transforms the traditional GZSL into a fully supervised learning task. Most generative GZSL methods use a single semantic attribute (each category can only correspond to a specific semantic attribute) and Gaussian noise to generate visual features, assuming a one-to-one correspondence between these visual features and single semantic attributes. However, in practice, there may be cases of attribute missingness in images, leading to visual features that lack certain attributes, thus failing to achieve a good mapping between semantic attributes and visual features. Therefore, visual features of the same class should have diverse semantic attributes. To address this issue, we propose a new method for enhancing semantic attributes, called “Semantic Attribute Interpolation Enhancement for Generalized Zero-Shot Learning (SAIE-GZSL).” This method uses interpolation to address the problem of semantic attribute missingness in real-world situations, thereby enhancing semantic diversity and generating more realistic and diverse visual features. We assessed the performance of the proposed model across four benchmark datasets, and the findings demonstrated substantial enhancements over current state-of-the-art methods, particularly in handling categories with severe attribute missingness in the datasets.
This article cannot obtain the latest cited-by information.
