计算机应用 ›› 2015, Vol. 35 ›› Issue (10): 2824-2827.DOI: 10.11772/j.issn.1001-9081.2015.10.2824

• 第十五届中国机器学习会议(CCML2015)论文 • 上一篇    下一篇

不完全鲁棒主成分分析的正则化方法及其在背景建模中的应用

史加荣, 郑秀云, 杨威   

  1. 西安建筑科技大学 理学院, 西安 710055
  • 收稿日期:2015-06-01 修回日期:2015-06-18 出版日期:2015-10-10 发布日期:2015-10-14
  • 通讯作者: 史加荣(1979-),男,山东东阿人,副教授,博士,CCF会员,主要研究方向:机器学习、模式识别,jiarongs3@163.com
  • 作者简介:郑秀云(1982-),女,山东日照人,讲师,博士,主要研究方向:最优化;杨威(1979-),女,辽宁抚顺人,副教授,博士,主要研究方向:信息融合、金融优化。
  • 基金资助:
    国家自然科学基金资助项目(61403298,11401457);陕西省自然科学基础研究计划项目(2014JQ8323,2014JQ1019);陕西省教育厅专项科研计划项目(2013JK0587)。

Regularized approach for incomplete robust principal component analysis and its applications in background modeling

SHI Jiarong, ZHENG Xiuyun, YANG Wei   

  1. School of Science, Xi'an University of Architecture and Technology, Xi'an Shaanxi 710055, China
  • Received:2015-06-01 Revised:2015-06-18 Online:2015-10-10 Published:2015-10-14

摘要: 针对现有的鲁棒主成分分析(RPCA)方法忽略序列数据的连续性及不完整性的情况,提出了一种低秩矩阵恢复模型——正则化不完全鲁棒主成分分析(RIRPCA)。首先基于序列数据连续性的度量函数建立了RIRPCA模型,即最小化矩阵核范数、L1范数和正则项的加权组合;然后使用增广拉格朗日乘子法来求解所提出的凸优化模型, 此算法具有良好的可扩展性和较低的计算复杂度;最后,将RIRPCA应用到视频背景建模中。实验结果表明,RIRPCA比矩阵补全和不完全RPCA等方法在恢复丢失元素和分离前景上具有优越性。

关键词: 鲁棒主成分分析, 低秩矩阵恢复, 背景建模, 核范数最小化, 增广拉格朗日乘子法

Abstract: Because the existing Robust Principal Component Analysis (RPCA) approaches do not consider the continuity and the incompletion of sequential data, one type of low-rank matrix recovery model, named Regularized Incomplete RPCA (RIRPCA), was proposed. First, the model of RIRPCA was constructed based on a metric function for evaluating the continuity, where the model minimized a weighted combination of the matrix nuclear norm, L1 norm and regularized term. Then, the augmented Lagrange multipliers algorithm was employed to solve the proposed convex optimization problem. This algorithm has good scalability and low computation complexity. Finally, RIRPCA was applied to the background modeling of videos. The experimental results demonstrate that the proposed method has the superiority of recovering missing entries and separating foreground over matrix completion and incomplete RPCA.

Key words: Robust Principal Component Analysis (RPCA), low-rank matrix recovery, background modeling, nuclear norm minimization, augmented Lagrange multiplier

中图分类号: