基于特征脸和多特征的人脸识别算法研究

【作者】 刘丽娜；

【导师】 乔谊正；

【作者基本信息】 山东大学 ， 模式识别与智能系统， 2006， 硕士

【摘要】 现代社会对于身份鉴别的准确性、安全性与实用性提出了更高的要求，传统的身份识别方法，例如：密码、IC卡等，已经不能很好满足这种要求，而人体丰富的生理和行为特征为此提供了一个可靠的解决方案，因而引起了国际学术界和企业界的广泛关注。 人脸识别技术作为生物特征识别技术的主流技术之一，是国内外研究和应用的热点。自动人脸识别系统有两个主要的环节：人脸检测和定位，人脸识别(鉴证)。本文首先介绍了人脸识别技术中的几个主要研究方向，然后针对人脸识别算法主成分分析法(Principal Component Analysis，简称PCA)，也就是特征脸方法进行了深入研究，引进了基于二阶特征的人脸识别，并提出了基于多特征的人脸识别算法。本文的主要工作可以概括如下： 研究了两种补偿光照变化的灰度预处理方法：灰度归一化和直方图均衡，采用特征脸法对二者做了识别对比实验。 针对传统的特征脸(PCA)方法中特征向量的选择问题，做了大量实验分析。分析了传统特征脸方法的优点与不足。 由于传统特征脸法对光照变化的适应性差，本文引进了基于二阶特征脸的识别算法。该方法通过“丢弃”传统特征脸方法得到的前数个反应光照信息的特征脸，克服光照干扰的影响。本文针对一、二阶特征脸权重系数的选择问题做了大量的实验分析。 上述两种识别方法均是基于整体特征的，在进行特征提取时图像中的所有像素给予了相等的地位，但是研究表明不同的脸部特征在识别中起着各不相同的作用。我们提出了基于多特征的人脸识别算法，即同时采用Eigenface、EigenUpper、EigenTzone进行识别。此方法将整体特征和局部分析结合起来，并且EigenUpper、EigenTzone的分割简单无需对人的眼睛、鼻子、嘴巴等进行精确定位。该方法有两种，一种是基于多特征和传统特征脸的，另一种是基于多特征和二阶特征脸的。更多还原

【Abstract】 An accurate automatic individual identification of person is critical to a wide range of application domains such as access control, electronic commerce, and welfare benefits disbursement. Traditional personal identification methods such as key word, IC card etc., however, suffer from a number of drawbacks and are unable to satisfy the security requirements of our highly inter-connected information society. Biometrics refers to an individual based on his/her physiological and/or behavioral traits. Naturely it is likely an identification panacea. In recent years, it is beginning to provide very powerful tools for the problems requiring positive identification.Human face recognition is one of the most reliable biometric technology. An automatic face recognition system has two important aspects, one is face detection and location, the other is face feature extraction and recognition or identification. In the paper we introduce several common used methods of the technologies on the two aspects. Then, the dissertation focuses on conventional face recognition algorithm -Principal Component Analysis (PCA) and its further improvement in which we introduce the second-order eigenface method first and then propose a method based on multiple feature combination. The main contributions of this thesis includes:Two methods of pre-processing aiming to deal with the variability in appearance are studied. They are gray normalization and gray statistic map equilibrium. We have also compared the methods using eigenface for face recognition experiments.We have conducted recognition experiments many times to study the problem of selecting how many eigenvectors is best for recognition.Because the conventional eigenface method is sensitive to the lighting variability in appearance, the second-order eigenface method is introduced. The method discards many eigenfaces that obtained by conventional eigenface method更多还原