Computer Engineering and Applications ›› 2016, Vol. 52 ›› Issue (18): 24-30.

Previous Articles     Next Articles

Local manifold reconstruction based semi-supervised multi-view image classification

DONG Xiwei   

  1. 1.School of Information Science and Technology, Jiujiang University, Jiujiang, Jiangxi 332005, China
    2.College of Automation, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
  • Online:2016-09-15 Published:2016-09-14

基于局部流形重构的半监督多视图图像分类

董西伟   

  1. 1.九江学院 信息科学与技术学院,江西 九江 332005
    2.南京邮电大学 自动化学院,南京 210003

Abstract: In order to improve the performance of classification by using information of multi-view features in semi-supervised scenario, firstly, by minimizing local reconstruction error of input feature vectors, proper edge weights can be learnt for the graph which is generated by using input feature vectors as vertexes of graph. Then, the edge weights are used for semi-supervised learning. This paper applies the manifold structure of input data which is captured by minimizing local reconstruction error of input feature vectors for semi-supervised learning, which is beneficial to improve the accuracy of label prediction in semi-supervised learning. For the using of multi-view features of training images, firstly, with the help of the technique of improved canonical correlation analysis, multi-view features with more discriminant information can be learnt, then the multi-view features with more discriminant information are used for image classification tasks by effectively fusing. Experimental results demonstrate that the proposed method can effectively perform discriminant tasks by well exploring discriminant information of multi-view feature representations of training samples in semi-supervised scenario.

Key words: image classification, label propagation, canonical correlation analysis, multi-view, semi-supervised

摘要: 为了在半监督情境下利用多视图特征中的信息提升分类性能,通过最小化输入特征向量的局部重构误差为以输入特征向量为顶点构建的图学习合适的边权重,将其用于半监督学习。通过将最小化输入特征向量的局部重构误差捕获到的输入数据的流形结构应用于半监督学习,有利于提升半监督学习中标签预测的准确性。对于训练样本图像的多视图特征的使用问题,借助于改进的典型相关分析技术学习更具鉴别性的多视图特征,将其有效融合并用于图像分类任务。实验结果表明,该方法能够在半监督情境下充分地挖掘训练样本的多视图特征表示的鉴别信息,有效地完成鉴别任务。

关键词: 图像分类, 标签传播, 典型相关分析, 多视图, 半监督