基于原型理论的相对属性学习

Prototype based relative attribute learning

对于表示学习的研究,合理的特征表示对系统性能的影响在许多分类问题中已经超越了分类器的作用,成为系统设计中最重要的组成部分.为此以心理学中认知科学的原型理论为基础,提出了一种新的特征表示方法.根据额外数据集中部分具有代表性的数据作为各类数据的代表,组成各类数据的原型数据集.通过学习数据与各原型之间的相对关系,得到衡量任意数据与原型数据集之间关系的函数,即等级函数(rank function).由此,任意数据都可以利用等级函数组来评价它们与原型之间的相对关系,以此作为数据的新特征表示用于分类任务.通过在MINST和Pubfig数据库上的实验验证可以看到,相比于灰度特征和属性特征,原型相对属性不但符合人类对于图像的认知,而且在识别性能上具有更高的精度.

According to the research on representation learning, a proper feature representation of data has a greater impact than classifiers on classification. It＇s almost become the most important part in system design. In this paper, based on prototype theorem in psychology, a new feature is proposed. Specifically, the prototype dataset is composed of representative data of extra datasets. Then, the rank functions are derived based on the relationship between the prototype dataset and any data set. Thus, any data could be represented via the rank functions and the values of the functions are their new features. The proposed method has been checked on the MINST database and Pubfig database. Compared with the gray-scale feature and attribute, the prototype based relative attribute is more reasonable and has better performance.