摘要
非负矩阵分解(NMF)是一种有效的子空间降维方法,凭借其可解释性在人脸识别方面有着较好的应用。而增量式非负矩阵分解(INMF)利用近似的原则将上一步迭代寻优的运算结果参与后续计算,有效改善了NMF算法运算规模随训练样本增多而不断增大的现象。文章提出的改进增量式非负矩阵分解算法(Improved Incremental Non-negative Matrix Factorization)在INMF的基础上进一步利用了新加入样本的类别信息,优化了算法中参与迭代的增量系数向量的初始化值,使目标函数在迭代求解时具有更快的收敛速度和全局寻优能力。通过在ORL和YALE人脸数据库上的实验表明,该算法在运算速度和识别率上均优于传统的NMF算法和INMF算法。
Non-negative matrix factorization is a useful method of subspace dimensionality reduction, and it has a good application on face recognition with its interpretability. The incremental non-negative matrix factorization uses the? previous result for later calculations based on approximation method, to avoid the problem that the computing cost grows rapidly while the training samples increasing. The improved incremental non-negative matrix factorization algorithm proposed in this paper improves the initial value of incremental coefficient vector by utilizing the class information of newly added samples, which makes the algorithm has faster convergence and better global search capacity in the process of iteration solution. Experimental results on ORL and YALE face databases show that the proposed method has certain advantages on convergence rate and recognition accuracy over NMF and INMF.
出处
《信息通信》
2016年第7期4-8,共5页
Information & Communications
基金
浙江省教育厅基金资助项目(Y201431023)
关键词
人脸识别
子空间降维
非负矩阵分解
增量学习
初始化
类别信息
Face recognition
Subspace dimensionality reduction
Non-negative matrix factorization
Incremental learning
Initialization
Class information
