中文摘要：

非负矩阵分解（NMF）算法可以提取图像的局部特征，然而NMF算法有两个主要缺点：a）当矩阵维数较大时，NMF算法非常耗时；b）当增加新的训练样本或类别时，NMF算法必须进行重复学习。为克服NMF算法这些缺点，提出了一种新的分块NMF算法（BNMF）。特别地，该方法还可用于增量学习。通过在FERET和CMUPIE人脸数据库上进行实验，结果表明该算法均优于NMF和PCA算法。

英文摘要：

Non-negative matrix factorization （NMF） can extract local features of images. However, NMF method has two main drawbacks. One shortcoming is that it is very time-consuming to deal with large matrices. The other is that it must implement repetitive learning, when the training samples or classes are incremental. In order to overcome these two limitations, this paper presented a novel block NMF （BNMF） method. In particular, it could be applied to incremental learning. Two face databases, namely FERET and CMU PIE face databases, were selected for evaluation. Comparing with NMF and PCA schemes, the proposed method gives superior results.