关键词: 零次学习, 知识图谱, 生成对抗网络, 图卷积神经网络, 图像识别

Abstract: Zero-shot learning (ZSL) is an important branch of transfer learning in the field of image recognition. The main learning method is to train the mapping relationship between the semantic attributes of the visible category and the visual attributes without using the unseen category, and use this mapping relationship to identify the unseen category samples, which is a hot spot in the current image recognition field. For the existing ZSL model, there remains the information asymmetry between the semantic attributes and the visual attributes, and the semantic information cannot well describe visual information, leading to the problem of domain shift. In the process of synthesizing unseen semantic attributes into visual attributes, part of the visual feature information was not synthesized, which affected the recognition accuracy. In order to solve the problem of the lack of unseen semantic features and synthesis of unseen visual features, this paper designed a ZSL model that combined attribute and graph semantic to improve the zero-shot learning’s accuracy. In the learning process of the model, the knowledge graph was employed to associate visual features, while considering the attribute connection among samples, the semantic information of the seen and unseen samples was enhanced, and the adversarial learning process was utilized to strengthen the synthesis of visual features. The method shows good experimental results through experiments on four typical data sets, and the model can synthesize more detailed visual features, and its performance is superior to the existing ZSL methods. 

Key words:  , zero-shot learning, knowledge graph, generative adversarial networks, graph convolution, image recognition