【作者】 周祥明；

【导师】 杨凡；

【作者基本信息】 浙江师范大学 ， 计算机软件与理论， 2009， 硕士

【摘要】 自动人脸识别(AFR)主要研究如何赋予计算机通过人脸辨别人物身份的能力,作为模式识别的一个研究领域,不仅具有非常重要的科学意义并且商业应用价值巨大。经过几十年的发展,AFR在受控环境和人工配合的情况下已经达到非常高的识别率,并出现了不少实际应用的商业系统。AFR系统主要有三个部分组成:人脸检测、人脸特征点定位、人脸特征提取和特征分类。人脸特征点定位作为AFR系统最重要的组成部分,它的定位精度在很大程度上影响着AFR系统的性能,同时它也是人脸表情识别和人脸三维建模的关键步骤。主动表观模型(Active Appearance Model,简称AAM)作为人脸特征点定位最主要和最有效的方法之一,被大量的学者研究并应用于实际的AFR系统中。本文对该方法在实际应用中遇到的问题进行了深入分析和研究,对其不足提出了相应的改进方法。主要的改进有以下三点:1)针对光照变化影响定位算法效果的问题,提出一种快速Gabor小波算法,用它提取输入图像的纹理信息用于AAM纹理建模和拟合计算。2)针对人脸某部分发生遮挡导致定位算法效果变差的问题,提出一种基于分块加权的AAM拟合算法。该方法先把人脸区域分成几个子区域,然后根据每个区域被遮挡的程度分配一个权重,在拟合过程中不断调整每个区域对应的权重从而到达消除遮挡干扰的目的。3)为了提高拟合算法的效率,提出基于多分辨率的AAM拟合算法。首先在低分辨率图像上进行拟合,因为该图像包含的纹理信息相对较少,所以拟合所需要的计算复杂度也相对较少,在该图像上得到一个相对接近人脸的模板位置后,然后在高分辨率图像上进行更精确的拟合。后本文采用Visual Studio 2005构建了一个基于AdaBoost人脸检测和AAM人脸特征点定位的演示系统,并将本文对特征点定位算法在鲁棒性上的改进应用于该系统中。更多还原

【Abstract】 Automatic Face Recognition (AFR) mainly studies on how to give a computer the capability of recognizing identification through people faces. As a research field of pattern recognition, it not only has very important scientific significance, but also has great commercial application values. After decades of development, AFR has reached a very high recognition rate in the controlled environment and artificial situation, lots of business systems applied practically. There are three main parts of an AFR system, face detection, facial feature points positioning, and facial feature extraction and feature classification. As the most important part of AFR system, the accuracy of facial feature points positioning extremely affects AFR systems’ performance, meanwhile it is the key step of facial expression recognition and 3D face modeling. Active appearance model (AAM), as one of the most important and efficient methods of facial feature points positioning, has been studied and applied to actual AFR systems by many scholars. In this paper, the problems of the method in practical application are encountered in the deep analysis and researches, the corresponding methods for its lack are pointed out as well.There are three main issues for improvement: 1) In terms of the accuracy of AAM under varied illumination, a fast Gabor wavelet algorithm is proposed to compute AAM Fitting and Texture Modeling; 2) For the problem of occlusion, a sub-block weighted AAM fitting algorithm is present, which first divides a region is into sub-blocks with different weights according to the proportion of occlusion. The weights of each sub-block are adjusted in the fitting process in order to eliminate occlusion influence; 3) To enhance the speed of the AAM fitting, a multi scale fitting strategy is put forward. Because lower scale image has less texture information, the speed of fitting is fast. Through the lower scale fitting which provides a better initial location, the fitting on high scale image is accelerated.Finally, we design a demo system of face detection using AdaBoost and facial feature localization using the improved AAM proposed in this paper. The system implemented under Visual Studio 2005 by appropriate designing its functional modules.更多还原