摘要
为了研究加权的类间离散度矩阵对人脸识别率的影响以及寻找影响识别率的因素,分别采用零空间投影法和非零空间投影法直接求解Fisher准则线性判别问题。零空间投影保留最具判别能力的样本类内离散度矩阵的零空间,非零空间投影保留样本类间离散度矩阵的非零空间,两种方法都很好地避免了高维小样本问题。通过对Yale人脸库的图像进行试验,结果表明类内离散度矩阵的零空间投影法要优于类间离散度矩阵的非零空间投影法,而加权值修正对识别率的影响并不明显,它的应用是有一定前提条件的。
The zero space projection and non-zero space projection are two kinds of method that can directly solve linear discriminative analysis problem.Both methods were adopted to study the effect of the weighted between-class scatter matrix on the face recognition,in order to explore the factors affecting the recognition rate.The zero space projection method reserves the zero space of within-class scatter matrix which has the most discriminate ability.The non-zero space projection method reserves the non-zero space of the between-class scatter matrix.The two methods can avoid the problem of high-dimensional small sample.Through the experiments tested in the Yale face database,we can see that the result of zero space projection is superior to the result of non-zero space projection.And the method of weighted between-class scatter matrix can not significantly improve recognition rate,so its application needs certain preconditions.
出处
《北京石油化工学院学报》
2011年第2期39-43,共5页
Journal of Beijing Institute of Petrochemical Technology
关键词
线性判别分析
样本类间离散度
样本类内离散度
人脸识别
linear discriminant analysis
between-class scatter
within-class scatter
face recognition
