Institutional Repository at Tsinghua University: 基于人脸配

Abstract: 人脸识别是模式识别和计算机视觉中最有活力和最有挑战性的领域之一，在很多领域有着非常广阔的应用前景。本文总结了几种经典的人脸识别方法：特征脸方法、基于LBP的方法、应用形状特征的方法以及AdaBoost方法。在此基础上，提出了一种比较新颖的人脸识别方法。该方法主要有以下三个特征： 第一，对于检测到的人脸，先用人脸配准技术定位人脸上的88个关键点，利用这些点将人脸图片卷绕到标准的形状上。通过配准-卷绕操作可使得识别系统对人的表情和姿态变化更具适应性。 第二，用一个比较新颖而有效的特征——Gabor直方图进行人脸特征提取。Gabor直方图可以同时反映Gabor特征的幅值和相角特征，能比较充分地利用Gabor特征对于不同人脸的区分性。 第三，将人脸识别这一多类分类问题看成是对样本的排序问题，将“排序”的思想引进到人脸识别中，并且用RankBoost的方法对排序过程进行优化。理论和实验两个方面都表明，用RankBoost进行人脸识别可以大大降低问题的复杂度，对人脸识别系统的性能有很大的提高。另外，我们认为这一方法也可用于解决其它的多类分类问题。
Face recognition is one of the most active and challenging problems in computer vision and pattern recognition field. In this paper, some classical face recognition approaches are summarized and a novel method is proposed. There are mainly three key contributions in this method: First of all, this method localizes 88 feature points of human face using face alignment technique and then warps the face image into a standard shape. This technique can enhance the robustness of face recognition method to facial expression and pose variations. Secondly, a novel and useful feature, Gabor Wavelet Histogram is extracted from human face image. Gabor Wavelet Histogram can reflect both magnitude and phase property of Gabor features, which can make full use of the discriminability of Gabor features. The last but the most important, this method regards face recognition, a classical multi-category classification problem, as a ranking problem. The idea is to consider ranking in face recognition and apply RankBoost in the procedure of feature selection. Both theorical and experimental results show that, RankBoost can reduce the complexity of face recognition problem, and significantly improve the system's performance. We argue that it can be applied to other multi-category problems too.