Research on Zero-Shot Learning Based on Generative Model

Abstract

Zero-shot learning is mainly applied in image classification tasks. Traditional classification methods use a large number of manually labeled data samples, which is time-consuming and inefficient. Zero-shot learning use semantic information to recognizes unknown images in the absence of labeled training data, saving labor costs. The mainstream method of early zero-shot learning, which only used known class data during training, was susceptible to domain drift and pivot point issues during testing, resulting in a decrease in model performance on the test set. In recent years, researchers have begun to use the generative model to solve the above problems. By generating the sample features of unknown classes to make up for the lack of visual data of unknown classes in training, the problem of imbalance between known and unknown class data has been alleviated. Firstly, introducing the concept of zero-shot learning briefly. Secondly, it briefly introduces traditional zero-shot learning and generalized zero-shot learning. Thirdly, two kinds of zero-shot learning methods based on the generative model are emphatically introduced. Then, the datasets and evaluation methods of zero-shot learning are briefly introduced, and the experimental results of the classical model are compared and analyzed. Finally, make a summary and outlook.