单训练样本条件下的人脸识别算法研究

由于主成分分析法和线性判别分析法等传统方法对单训练样本的识别能力弱,甚至直接失效。本文提出了二维小波变换与矩阵的最大间距准则或矩阵的线性判别分析相融合的人脸特征提取算法。即首先将原图像进行三层二维小波变换,然后对每层的近似分量分别进行最大间距准则或线性判别分析处理,最后用欧氏距离判别。在ORL人脸数据库上取得的实验结果表明,本文提出的算法能够提高单训练样本条件下的人脸识别率,同时也满足实时性要求。

Face recognition from a single training sample

The recognition ability of conventional algorithms like principal component analysis（PCA） andlinear discriminant anal-ysis （LDA）is usually poor for a single training sample. The hybrid method of the two-dimensional wavelet transform and maximum margin criterion（MMC） or linear discriminant analysis is proposed. That is, a source image is firstly processed by the three layers two-dimensional wavelet transform. Secondly, MMC or LDA is employed to handle the approximation coefficients. The Euclidean dis-tance is used to distinguishthecategory. The experimental resultsconducted on the ORL database illustrate that the proposed method can improve the recognition rate and meet the real-time requirements for a single training sample.