电子学报 ›› 2015, Vol. 43 ›› Issue (3): 440-446.DOI: 10.3969/j.issn.0372-2112.2015.03.004

• 学术论文 • 上一篇    下一篇

基于低秩分解的联合动态稀疏表示多观测样本分类算法

胡正平, 高红霄, 赵淑欢   

  1. 燕山大学信息科学与工程学院, 河北秦皇岛 066004
  • 收稿日期:2013-08-30 修回日期:2014-04-09 出版日期:2015-03-25
    • 作者简介:
    • 胡正平 男,1970年生于四川仪陇县,教授,博士导师,燕山大学通信电子工程系主任,1996年于燕山大学无线电专业获得学士学位,并获得推荐研究生资格,1999获得电路与系统硕士学位,2007年于哈尔滨工业大学获得信息与通信工程专业博士学位,目前为中国电子学会高级会员,中国图像图形学会高级会员,目前研究方向为稀疏模式识别. E-mail:hzp@ysu.edu.cn 高红霄 女,1987年生于河北石家庄,燕山大学信息科学与工程学院通信与信息系统专业,主要研究方向为多观测样本分类. E-mai:970353066@qq.com
    • 基金资助:
    • 国家自然科学基金 (No.61071199)

Multiple Observation Sets Classification Algorithm Based on Joint Dynamic Sparse Representation of Low-Rank Decomposition

HU Zheng-ping, GAO Hong-xiao, ZHAO Shu-huan   

  1. School of Information Science and Engineering, Yanshan University, Qinhuangdao, Hebei 066004, China
  • Received:2013-08-30 Revised:2014-04-09 Online:2015-03-25 Published:2015-03-25
    • Supported by:
    • National Natural Science Foundation of China (No.61071199)

摘要:

通过互联网易获得同一对象的多个无约束的观测样本,针对如何解决无约束观测样本带来的识别困难及充分利用多观测样本数据信息提高其分类性能问题,提出基于低秩分解的联合动态稀疏表示多观测样本分类算法.该算法首先寻找到一组最佳的图像变换域,使得变换图像可以分解成一个低秩矩阵和一个相关的稀疏误差矩阵;然后对低秩矩阵和稀疏误差矩阵分别进行联合动态稀疏表示,以便充分利用类级的相关性和原子级的差异性,即使多观测样本的稀疏表示向量在类级别上分享相同的稀疏模型,而在原子级上采用不同的稀疏模型;最后利用总的稀疏重建误差进行类别判决.在CMU-PIE人脸数据库、ETH-80物体识别数据库、USPS手写体数字数据库和UMIST人脸数据库上进行对比实验,实验结果表明本方法的优越性.

关键词: 模式识别, 多观测样本分类, 低秩矩阵恢复, 联合动态稀疏表示

Abstract:

Multiple unconstrained observations of the same object can be easily accessed by the Internet, with regard to overcoming the identification-difficult of the unconstrained samples.Moreover, to exploit the information of multiple observation sets to improve the classification performance, a multiple observation sets classification algorithm based on joint dynamic spare representation of low-rank decomposition is presented.First of all, we need find the best set of image transform domain, which decomposes the data matrix into a low-rank matrix and an associated sparse error matrix.Secondly, the low-rank matrix and sparse error matrix is represented by joint dynamic sparsity respectively, in order to make full use of the correlation of the class-level and the differences of the atom-level, i.e, the sparse representation vectors for the multiple observations can share the same class-level sparsity pattern while their atom-level sparsity patterns may be distinct.Finally, we compare the classification results with the total sparse reconstruction errors.Three comparative experiments are conducted on CMU-PIE face dataset, ETH-80 object recognition dataset, USPS handwritten digit dataset, and UMIST face dataset, and the results demonstrate the superiority of the proposed algorithm.

Key words: pattern recognition, multiple observation sets classification, low-rank matrix recovery, joint dynamic sparse representation

中图分类号: