一种面向零样本图像分类的交互式类属性构建方法
An Interactive Class Attribute Construction Method For Zero-shot Image Classification
-
摘要: 零样本图像分类可以解决训练数据类和测试数据类不相交时的图像分类问题, 人类标注属性是一种常用的实现零样本图像分类的辅助知识. 设计一个合适的类属性矩阵需要领域专家全面分析各个类别的异同点, 定义各个类具有的属性, 这是个繁琐且无任何指导的过程. 为了辅助专家用户设计用于零样本图像分类的类属性矩阵, 本文提出了一种交互式构建方法. 首先通过一种基于概念的深度学习可解释性方法, 在训练集数据中提取属性; 其次通过多视图协作的交互方式, 探索和分析已提取属性的重要性; 再次系统提供了全局和局部两种方式, 辅助用户设计测试集数据类别的属性值. 最后在零样本学习的基准数据集上进行案例分析, 实验表明该系统可以帮助专家用户高效和便捷地完成类属性构建工作.Abstract: Zero-shot image classification can solve the image classification problem when the training data classes and test data classes are disjoint. Human annotation attributes are a commonly used auxiliary knowledge to achieve zero-shot image classification. Designing a suitable class attribute matrix requires comprehensive analysis by domain experts. It is a tedious and unguided process to define the similarities and differences of each category and define the attributes of each category. In order to assist expert users in designing a class attribute matrix for zero-shot image classification, this paper proposes an interactive construction method. First Through a concept-based deep learning interpretability method, attributes are extracted from the training set data; secondly, the importance of the extracted attributes is explored and analyzed through the interactive method of multi-view collaboration; thirdly, the system provides global and local ways to assist users in designing the attribute values of the test set data categories. Finally, a case analysis is carried out on the benchmark dataset of zero-shot learning. The experiments show that the system can help expert users to complete the construction of class attributes efficiently and conveniently.