基于图像分析的人脸比对技术研究

【作者】 王峰；

【导师】 任明武；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2012， 硕士

【摘要】 基于图像分析的人脸比对技术研究主要是面向ATM机环境的自动监控、报警和身份验证的问题提出来的,通过人脸异常的判断和人脸比对两个步骤实现以上的功能。针对正面人脸特征比较多,而且容易提取,本文主要研究了如何对正面人脸进行比对研究。在研究和学习前人理论和方法的基础上,提出了一种基于加权分块区域的正面人脸比对方法。本文的主要研究工作如下：(1)利用Adaboost算法构建了人眼检测和嘴巴检测的分类器,并对训练过程中矩形特征的数量提出了一些优化的方法和改进的措施。通过检测人眼和嘴巴是否缺失来判断是否存在人脸异常的情况。(2)对于前面分析留下的正常人脸图像存在的一些瑕疵,提出了一些校正的方法。并且规范化处理了所有的人脸图像,为人脸比对提供了较好的素材。(3)提出了基于加权分块区域的正面人脸比对方法。该方法首先利用主动形状模型(ASM)的方法实现人脸特征点的自动标定。基于人脸面部的先验知识,面部几何位置特征存在稳定性。然后对特征点分成5块区域,构造人脸个性特征参数,分析各个区域的权重,在此基础上进行加权分块区域的人脸比对。最后实验论证了该方法对于正面人脸具有比较好的比对性能。更多还原

【Abstract】 Face matching research based on image analysis issues to the ATM machine environment for automatic monitoring, alarm and authentication by two steps to achieve the above functions which are face abnormalities judgments and face matching.As frontal face include more features, and easy to extract, this paper mainly studies how to match the frontal face. Based on researching and learning the previous theory and methods, a face matching algorithm with block based on the weighted region is presented.What had been mainly done in this paper is as follows:1. Adaboost algorithm have been used to build eye detection and mouth detection classifier, furthermore, some optimized methods and improved measures about the number of rectangular features in the training process are presented. We determine whether there is abnormal situation in the face image by detecting whether the absence of eye and mouth.2. Some correction methods are presented for the remaining normal images which have some flaws. Moreover, all defective face images have been corrected, which provides good material for face matching.3. A face matching algorithm with block based on the weighted region is presented. Firstly, facial feature points are labeled automatically by using Active Shape Models (ASM). According to facial prior knowledge, the features of the facial geometry information are stable. Then the feature points are divided into five regions, and facial personality features are constructed, and the weight of each region is analyzed, and face matching with block based on the weighted region have been carry on by these steps. At last, this method has better performance for frontal face matching demonstrated through experiments.更多还原