计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (15): 153-160.DOI: 10.3778/j.issn.1002-8331.1901-0333

• 模式识别与人工智能 • 上一篇    下一篇

共有结构假设下流形正则图的零样本分类方法

马丽红,谭学仕   

  1. 华南理工大学 电子与信息学院,广州 510641
  • 出版日期:2019-08-01 发布日期:2019-07-26

Zero-Shot Classification with Manifold Regularization Graph Based on Common Structure Assumption

MA Lihong, TAN Xueshi   

  1. School of Electronic and Information Engineering, South China University of Technology, Guangzhou 510641, China
  • Online:2019-08-01 Published:2019-07-26

摘要: 零样本学习(Zero-Shot Learning,ZSL)利用视觉和语义特征关联模型进行可鉴别知识迁移,但视觉和语义数据不是简单的对应关系,难以直接建立映射函数。提出一种局部敏感双字典方法,主要贡献有两点:(1)双字典方法。视觉-语义的单字典映射缺乏直接关联的共有变量,提出双字典方法为视觉和语义添加一个共有结构的描述字典,从而构造更合理的视觉-语义关联通道。(2)局部敏感的流形保持方法。在双字典学习中,局部结构信息的描述是关键点,通过构造流形结构图来定义局部敏感约束项,对字典学习和局部流形保持进行联合优化。在AwA和CUB数据集上的实验结果表明,该方法在分类准确率上优于对比算法。

关键词: 零样本学习, 知识迁移, 双字典, 共有结构, 局部敏感

Abstract: Zero-Shot Learning(ZSL) has utilized association model of visual and semantic features to transfer discriminative knowledge. But visual and semantic data is not a simple correspondence, it’s difficult to directly establish the mapping. A locality sensitive double dictionary method is proposed, which has two main contributions:(1) Double dictionary method. Visual-semantic single dictionary mapping lacks direct associated common variables, a double dictionary method is proposed to add a descriptive dictionary of common structure for visual and semantic features, a more reasonable visual-semantic association channel is constructed. (2) Locality sensitive manifold preserving method. In the double dictionary learning, the description of local structure information is vital. A manifold structure graph is constructed to define locality sensitive regularization, and to jointly optimize the dictionary learning and local manifold preserving. The experimental results on AwA and CUB datasets show the proposed method outperforms the compared algorithms in accuracy.

Key words: zero-shot learning, knowledge transfer, double dictionary, common structure, locality sensitive