Semantic Feedback for Generalized Zero-Shot Learning
Luojing Zhang, Guangjian Zhang · 2024
The goal of Generalized Zero-Shot Learning (GZSL) methods is to classify unseen classes during testing, where no available data is provided during training. Unlike zero-shot recognition, test samples of GZSL include both seen and unseen classes. Current feature generation methods are based on generative adversarial networks, which synthesize visual features for unseen classes using their semantic descriptors. However, these methods still struggle to address the bias between seen and unseen classes. Therefore, this paper proposes a Semantic Feedback (SF) module based on generative adversarial models to mitigate the bias between seen and unseen classes. Specifically, we introduce a semantic feedback module with a feedback mechanism that iteratively improves the generated features during the training and feature synthesis stages. The synthesized visual features, along with their corresponding latent features, are then refined into discriminative features for the classification process to reduce class ambiguity. Competitive results obtained on four popular benchmark datasets demonstrate the significant potential and superiority of this proposed method.