计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

面向分类的增量字典学习算法

张志武 1a,荆晓远 1b,2,吴飞 1b   

  1. (1.南京邮电大学 a.计算机学院; b.自动化学院,南京 210023; 2.武汉大学 软件工程国家重点实验室,武汉 430072)
  • 收稿日期:2016-08-23 出版日期:2017-10-15 发布日期:2017-10-15
  • 作者简介:张志武(1981—),男,博士研究生,主研方向为模式识别、机器学习;荆晓远(通信作者),教授、博士、博士生导师;吴飞,博士。
  • 基金项目:
    国家自然科学基金(61073113);江苏省普通高校研究生科研创新计划项目(CXZZ12_0478)。

Classification-oriented Incremental Dictionary Learning Algorithm

ZHANG Zhiwu 1a,JING Xiaoyuan 1b,2,WU Fei 1b   

  1. (1a.School of Computer; 1b.School of Automation,Nanjing University of Posts and Telecommunications, Nanjing 210023,China; 2.State Key Laboratory of Software Engineering,Wuhan University,Wuhan 430072,China)
  • Received:2016-08-23 Online:2017-10-15 Published:2017-10-15

摘要: 针对面向分类的传统字典学习方法在大数据集上批量学习时计算代价较高的问题,提出一种类特定的增量式字典学习算法。该算法在初始训练集上进行类特定的字典学习得到初始字典,通过增量数据集选取增量字典原子初始值。根据不能在初始字典上稀疏表示且互信息最大的原则,从增量样本集中选取若干样本作为增量字典原子的初始值。在保持原有字典原子不变的情况下,迭代更新编码系数和增量字典原子,直至收敛得到新的字典。利用稀疏表示分类器,在Eclipse数据集和ORL人脸图像数据库上的实验结果验证了该算法的分类有效性和计算代价上的优越性。

关键词: 增量学习, 字典学习, 类特定字典, 稀疏编码, 稀疏表示分类

Abstract: Aiming at the problem that the computation cost of the traditional classification-oriented dictionary learning algorithms is too expensive on big datasets,this paper proposes a novel classification-oriented incremental dictionary learning algorithm.In the algorithm,the class-specific dictionary learning is conducted on the initial training set to obtain the initial dictionary.And the initial values of the incremental dictionary atoms are selected on the incremental data set.Based on the principle that the samples cannot be sparsely represented by old atoms and have the maximum mutual information,some samples are selected as the initial value of the incremental dictionary atoms.Keeping the original dictionary unchanged,the coding coefficient and the incremental dictionary atoms are updated iteratively until the convergence is realized and the new dictionary is obtained.Sparse representation classifier is used as classifier in experiments.Experimental results on the Eclipse software defect dataset and ORL face image database show that the proposed algorithm is effective in classification and has superiority in computational cost.

Key words: incremental learning, dictionary learning, class-specific dictionary, sparse coding, sparse representation classification

中图分类号: