计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (23): 193-199.DOI: 10.3778/j.issn.1002-8331.2001-0041

• 模式识别与人工智能 • 上一篇    下一篇

融合数据增强的迁移字典学习

王子儒,李振民   

  1. 中南大学 自动化学院,长沙 410083
  • 出版日期:2021-12-01 发布日期:2021-12-02

Transferable Dictionary Learning Fused Data Augmentation

WANG Ziru, LI Zhenmin   

  1. School of Automation, Central South University, Changsha 410083, China
  • Online:2021-12-01 Published:2021-12-02

摘要:

提出利用迁移字典解决复杂行为数据集标签样本不足的问题。所提出的方法使用简单行为作为源域,来辅助识别由一系列简单行为组成的复杂行为。通过稠密轨迹提取视频的低级特征,利用字典学习从简单行为和复杂行为的低级特征中分别获得相应的稀疏表示,并利用简单行为的稀疏表示通过迁移矩阵改善复杂行为的稀疏表示。因此,即使在复杂行为标签样本较少的情况下,迁移字典也能够获得更有效的高级特征。同时,利用GAN在特征层面上进行数据增强,帮助学习表征能力更强的字典。提出的方法在UCF101和HMDB51两个数据上进行了实验,在小样本量的情况下获得了比现有方法更好的识别结果,证明了方法的有效性。

关键词: 复杂行为识别, 迁移字典, 特征增强

Abstract:

A transferable dictionary method is proposed to solve the problem that insufficient label samples in complex behavior dataset. The proposed method uses simple action as the source domain to assist in identifying complex action composed of a series of simple actions. The low-level features of video are extracted by dense trajectory, and then the sparse representation of simple action and complex action are obtained by dictionary learning, and the sparse representation of simple action is used to improve the sparse representation of complex action by transformation matrix. Therefore, even in the case of fewer complex action labeled data, the transferable dictionary can obtain more efficient features. At the same time, GAN is used to data augmentation at the feature level, which helps to learn the dictionary with stronger representation ability. The proposed method is tested on UCF101 and HMDB51 dataset, and obtains better recognition results than the existing method in the case of small sample size, which proves the effectiveness of the method.

Key words: complex action recognition, transferable dictionary, feature augmentation