摘要
当源域和目的域数据分布不同时,大多数机器学习方法的性能会降低。为了解决这一问题,基于域适应的思想,提出了一种新的人脸识别方法。首先计算源域样本的相对权值,删除与目的域样本相差很大的样本,降低两域之间的差异性;然后采用基于正规化的Bregman divergence获得公共子空间,获得两域之间的共性;最后利用目的域样本目标化源域样本,充分利用目的域的特有信息。在此基础上建立的分类模型能够充分利用两域之间的共性和目的域的特性,实现对目的域的准确分类。为了评估方法的性能,在多个数据集上进行测试实验。实验结果证明,该方法的性能与其他几种方法相比均有所提高。
The performance of most machine learning technologies becomes poorly when the training data and test data have different distributions. In order to solve the problem, this paper presented a method based on domain adaptation. The basic idea of the method was to calculate relative importance weight for the training samples and then deleted some training samples which were wildly different from test data. As a result,it reduced the disparity of two domains. Then it learned a common subspace by Bregman divergence based regnlarization and getting the commonality between the source and target domains. The discriminative model based on the common subspace could make full use of the commonality between two domains and particular knowledge of the target domain. It evaluated the proposed method on multiple datasets. The results show that the method has better performance than other existing methods.
出处
《计算机应用研究》
CSCD
北大核心
2017年第6期1881-1884,共4页
Application Research of Computers
基金
国家自然科学基金委员会-山西省人民政府煤基低碳联合基金资助项目(U1510115)
中国博士后科学基金特别资助项目(2013T60574)
作者简介
周军娜(1990-),女,江苏徐州人,硕士研究生,主要研究方向为图像处理、目标跟踪(1260166339@qq.com);
陈伟(1978-),男,教授,博士(后),主要研究方向为无线传感器、图像处理;
王珂(1978-),男,副教授,博士,主要研究方向为无线传感器、图像处理;
汤镇宇(1990-),男,硕士研究生,主要研究方向为图像处理.