Enhancing Semantic-Consistent Features and Transforming Discriminative Features for Generalized Zero-Shot Classifications

Generalized zero-shot learning (GZSL) aims to classify classes that do not appear during training. Recent state-of-the-art approaches rely on generative models, which use correlating semantic embeddings synthesize unseen visual features; however, these ignore the and relevance, features synthesized by models represent their semantics well. Although existing GZSL methods based model disentanglement consider consistency between only in training phase feature synthesis classification phases. The absence of such constraints may lead an unrepresentative with respect semantics, are modally well aligned, thus causing bias features. Therefore, approach for is proposed enhance semantic-consistent discriminative transformation (ESTD-GZSL). method can at all stages GZSL. A decoder module first added VAE map synthetic real corresponding embeddings. This regularization allows synthesizing a more representative representation, better semantics. Then, decomposed output transformed into enhanced used reduce ambiguity categories. experimental results show our achieves competitive four benchmark datasets (AWA2, CUB, FLO, APY)