基于对称非负矩阵分解的人脸识别算法

FACE RECOGNITION ALGORITHM BASED ON SYMMETRICAL NMF

提出一种利用镜像偶特征改造NMF(Non-negative Matrix Factorization)基图像并进行局部特征提取的人脸识别算法。首先获取镜像偶特征并进行二次Haar小波分解,得到重构人脸样本图像。然后利用NMF分解得到一组在垂直方向对称的基图像,由它们组成基矩阵并对它们正交规范化。改造后的基图像符合人脸对称的生理特性,使得NMF基矩阵更加适用于人脸特征提取。在含有姿态变化和不均匀光照样本的Yale人脸数据库上取得了较好的识别效果。

This paper proposes an algorithm for face recognition by reforming the base image of NMF with even mirror-like features and extracting the local feature. First, the even mirror-like feature is obtained and followed by the secondary Haar wavelet decomposition, and the reconstructed face sample image is derived. Then, by employing NMF method a set of basic images symmetrical in vertical direction are obtained, and they are used to form the base matrix and to be performed the orthogonal normalisation. The transformed base image meets the physiological characteristic of facial symmetry, this makes the NMF base matrix more suitable for face feature extraction. The algorithm achieves pretty good recognition effect on Yale face database with the samples of posture changes and uneven illumination.