摘要:零样本学习旨在通过运用已学到的已知类知识去认知未知类.近年来，“数据+知识驱动”已经成为当下的新潮流，而在计算机视觉领域内的零样本任务中，“知识”本身却缺乏统一明确的定义.针对这种情况，尝试从知识的角度出发，梳理了本领域内“知识”这一概念所覆盖的范畴，共划分为初级知识、抽象知识以及外部知识.基于前面对知识的定义和划分，梳理了当前的零样本学习（主要是图像分类任务的模型）工作，分为基于初级知识的零样本模型、基于抽象知识的零样本模型以及引入外部知识的零样本模型.还对领域内存在的域偏移和枢纽点问题进行了阐述，并基于问题对现有工作进行了总结归纳.最后总结了目前常用的图像分类任务的数据集和知识库、图像分类实验评估标准以及代表性的模型实验结果，并对未来的工作进行了展望.



Abstract:Zero-shot learning aims to recognize the unseen classes by using the knowledge of the seen classes that has been learned. In recent years, ‘knowledge+data driven’ has become a new trend but lacking of unified definition of ‘knowledge’ in the current zero-shot tasks of computer vision. This study tries to define the ‘knowledge’ in this field and divided it into three categories, which are primary knowledge, abstract knowledge, and external knowledge. In addition, based on the definition and classification of knowledge, the current works on zero-shot learning (mainly in image classification task) are sorted out, they are divided into zero-shot models based on primary knowledge, zero-shot models based on abstract knowledge, and zero-shot models based on external knowledge. This study also introduces the problems which are domain shift and hubness in this field, and further summarizes existing works based on the problems. Finally, the paper summarizes the datasets and knowledge bases that commonly used in image classification tasks, the evaluation criteria of image classification experiment and the experimental results of representative models. The future works are also summarized and prospected.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/6146