基于描述特征的人脸识别研究

针对基于主成分分析识别人脸存在计算复杂、不能准确地估计训练图像的协方差矩阵等问题,提出了一种基于描述特征的人脸识别算法(Expressive feature face recognitionalgorithm, EFFRA).该算法用训练图像的右奇异向量代替PCA求解的子空间的基向量,避免了将人脸图像转换成图像向量,明显降低了计算复杂性.进一步研究发现,EFFRA提取的每一个主成分向量中含有冗余,在此基础上,利用PCA实现了EFFRA的简化算法(MEFFRA),在ORL和Essex数据库上的实验结果表明,EFFRA及MEFFRA明显优于特征脸算法,MEFFRA的识别精度略好于EFFRA,但明显减少了对存储空间的需求.

Face Recognition Based on Expressive Features

The principal component analysis (PCA) faces the problem of high computation complexity and inaccurate estimated covariance matrix from training face images for face recognition. The expressive feature face recognition algorithm (EFFRA) is proposed. In EFFRA, the subspace basic vector extracted by PCA is substituted by the right singular vectors of training images, so that the transformation from the images to image vectors is avoided. Hence the computation is simplified significantly. Further analysis shows that each principal component vector extracted by EFFRA still contains redundancy. Based on this result, a modified EFFRA (MEFFRA) is presented by combining the EFFRA and PCA. The results based on ORL and Essex database show that both EFFRA and MEFFRA are superior to eigenfaces, recognition ability of MEFFRA is no less than EFFRA with a much smaller storage space compared with EFFRA.