基于嵌入对比学习的广义零样本预分类模型

Generalized zero-shot pre-classification model based on embedding contrastive learning

广义零样本学习,需要结合视觉和语义信息,识别可见和不可见类。本文提出基于嵌入对比学习的广义零样本预分类模型。该模型利用特殊的自编码器获取多模态潜在空间,并利用对比学习,对齐视觉和语义特征并进行优化。通过这种方式,实现更好的类内相似性和预测精度。实验证明,该模型在四个数据集上取得了良好效果。

Generalized zero-shot learning combines visual and semantic information to identify visible and invisible classes.A generalized zero-shot pre-classification model based on embedding contrastive learning is proposed.This model employs a specialized autoencoder to obtain a multimodal latent space and utilizes contrastive learning to align and optimize visual and semantic features,achieving better within-class similarity and prediction accuracy.Experiments demonstrate that the proposed model achieves promising results on four datasets.