Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (5): 812-824.DOI: 10.3778/j.issn.1673-9418.2010092

• Surveys and Frontiers • Previous Articles     Next Articles

Survey of Zero-Shot Image Classification

LIU Jingyi, SHI Caijuan, TU Dongjing, LIU Shuai   

  1. 1. College of Artificial Intelligence, North China University of Science and Technology, Tangshan, Hebei 063210, China
    2. College of Electronics and Information Engineering, Shenzhen University, Shenzhen, Guangdong 518060, China
  • Online:2021-05-01 Published:2021-04-30

零样本图像分类综述

刘靖祎史彩娟涂冬景刘帅   

  1. 1. 华北理工大学 人工智能学院,河北 唐山 063210
    2. 深圳大学 电子与信息工程学院,广东 深圳 518060

Abstract:

It is time-consuming and laborious to manually label a large number of samples, and samples from some rare classes are difficult to obtain. Therefore, the zero-shot image classification has become a research hotspot in the computer vision field. Firstly, the zero-shot learning, including direct push zero-shot learning and inductive zero-shot learning, is introduced briefly. Secondly, the space embedding zero-shot image classification methods and the generative model based zero-shot image classification methods with their subclass methods are introduced emphatically. Meanwhile, the mechanism, advantages and disadvantages, and application scenarios of these methods are analyzed and summarized. Thirdly, the main datasets and main evaluation criteria for zero-shot image classification are briefly introduced, and the performance of typical zero-shot image classification methods is compared. Then, the problems such as domain drift, hubness and semantic gap and the corresponding solutions are pointed out. Finally, the future development trends and research hotspots of zero-shot image classification are discussed, such as the accurate location of discriminative region, visual features of high-quality unseen class, generalized zero-shot image classification, etc.

Key words: zero-shot learning, zero-shot image classification, embedding space, generative?model, deep learning

摘要:

面对人工标注大量样本费时费力,一些稀有类别样本难于获取等问题,零样本图像分类成为计算机视觉领域的一个研究热点。首先,对零样本学习,包括直推式零样本学习和归纳式零样本学习进行了简单介绍;其次,重点介绍了基于空间嵌入零样本图像分类方法和基于生成模型零样本图像分类方法以及它们的子类方法,并对这些方法的机制、优缺点和适用场景等进行了分析和总结;然后,简单介绍了零样本图像分类常用数据集和评估方法,并对典型零样本图像分类方法进行了性能比较;接着,指出了现有零样本图像分类中存在的领域漂移、枢纽点和语义鸿沟等问题及相应的解决思路;最后,对零样本图像分类未来发展趋势和研究热点,如判别性区域的准确定位、生成高质量不可见类视觉特征、广义零样本图像分类等进行了探讨。

关键词: 零样本学习, 零样本图像分类, 嵌入空间, 生成模型, 深度学习