摘要
对于单训练样本人脸识别,基于每人多个训练样本的传统人脸识别算法效果均不太理想。尤其是基于Fisher线性鉴别准则的一些方法,由于类内散布矩阵为零矩阵,根本无法进行识别。提出一种新的样本扩充方法,即泛滑动窗法。采用"大窗口、小步长"的机制进行窗口图像采集和样本扩充,不仅增加了训练样本,而且充分保持和强化了原始样本模式固有的类内和类间信息。然后,使用二维线性鉴别分析方法(2DLDA)对上面获得的窗口图像进行特征抽取。在ORL国际标准人脸库上进行的实验证实了所提算法的可行性和有效性。
For face recognition with single training sample per person, the conventional face recognition methods which work with many training samples do not function well. Especially, a number of methods based on Fisher linear discrimination criterion can not work because the within-class scatter matrix is a matrix with all elements being zero. To solve this problem, a new sample augment method, called generalized slide window, was proposed. In order to effectively maintain and strengthen the within-class and between-class information, the rule, "big window, small step", was adopted to produce a set of window images for each training image. Then, two-dimensional Fisher linear discrimination analysis was performed on the window images obtained. The experimental results on ORL face database confirm that the proposed method is feasible and effective in face recognition with single training sample per person.
出处
《计算机应用》
CSCD
北大核心
2007年第11期2793-2796,2807,共5页
journal of Computer Applications
基金
国家自然科学基金资助项目(60472060)
江苏省高校自然科学基金资助项目(05KJB520152)
江苏省博士后科研资助计划项目
关键词
单样本
泛滑动窗
特征抽取
人脸识别
single training sample
generalized slide window
feature extraction
face recognition