基于二维线性判别分析的人脸识别算法的研究.docVIP

  基于二维线性判别分析的人脸识别算法研 究# 王丽莉，周亚同，王宝珠** 5 10 （河北工业大学信息工程学院） 摘要：在现有成熟的人脸识别算法中，主成份分析法（PCA）侧重于压缩，不利于分类，线 性判别分析（LDA）算法虽适合于分类，但存在小样本的问题。PCA+LDA 算法是先利用 PCA 技术对原图像进行降维处理，使得类内离散度矩阵非奇异，之后再用 LDA 的方法进行 特征提取，但在利用 PCA 降维的同时会丢失部分图像信息，影响后续的分类。为此，本文 研究了双向二维线性判别分析 2DLDA 的方法，该方法是直接从人脸图像矩阵求得类内和类 间离散度矩阵，并不需要将人脸图像矩阵转化为向量，不仅解决了小样本问题，也减少了计 算量，实验结果也表明 2DLDA 算法相对于 PCA、PCA+LDA 算法具有一定的优势 关键词：人脸识别；主成分分析；线性判别分析；PCA+LDA；二维线性判别分析 中图分类号：TP391.9 15 study on face recognition algorithm based on two-dimensional linear discriminant analysis WANG Lili, ZHOU Yatong, WANG Baozhu (School of Information Engineering, Hebei University of Technology) 20 25 30 35 Abstract: In the existing mature face recognition algorithms, the principal component analysis (PCA) method focuses on compression, is not conducive to the classification. The linear discriminant analysis (LDA) algorithm, although suitable for classification, but there are problems of small samples. PCA + LDA algorithm use PCA to reduce the dimension of the original image firstly, and to make the class scatter matrix nonsingular, then use LDA method for feature extraction. However, this method will also lose some image information when using PCA to reduce dimension, and is not good for the subsequent classification. To solve this problem. This paper studied the two-dimensional linear discriminant analysis method(2DLDA).The method obtaineed the within and between-class scatter matrix from the face image matrices directly, and does not need to change face images into vector matrix, not only solved the small sample size problem, but also reduced the amount of computation. Experimental results show that 2DLDA algorithm has certain advantages compared to the PCA, PCA + LDA algorithm. Key words: Face Recognition; PCA; LDA; PCA+LDA; 2DLDA 0 引言 随着社会对安全系统的迫切需要，人脸识别技术不仅可以应用到国家的安全方面，还可 以应用到军事和公共安全等方面。虽然人类可以毫不费力的辨别成千上万张人脸，但是利用 计算机实现完全自动的人脸识别还是非常困难的。人脸的表情丰富多彩，并且随着时间变化， 人脸的面部特征也会相应的发生一些变化，加上人脸图像会受到光照、角度以及成像距离等 方面的影响，这些都使得人脸识别成为一项极具挑战性的研究课题，具有很高的研究价值。 基金项目：河北省自