【作者】 王新竹；

【导师】 田彦涛；

【作者基本信息】 吉林大学 ， 控制理论与控制工程， 2012， 硕士

【摘要】 人脸表情识别涉及到人工智能、模式识别、图像处理、心理学等诸多领域，是研究人机交互的一个重要方向。随着计算机运行速度的不断提高，人们也越发关注基于动态图像序列的人脸表情识别的研究。基于动态图像序列的人脸表情识别不同以往的基于静态图像的表情识别研究。基于动态图像序列提取出的表情特征较静态特征含有更多表征表情的静态和动态信息，静态信息主要体现在每幅图像中的表情特征提取，而动态信息则是多个图像集合而成的整体表情变化方向。本论文的研究内容概述如下：1．首先针对基于Candide3三维人脸模型的头部姿态及面部表情跟踪算法进行了研究，设计了一种半自动的人脸模型匹配方法，用于跟踪算法中的模型参数初始化和形状无关纹理合时所需要的样本生成。在研究了原跟踪算法的基础上，根据存在的不足进行了跟踪算法流程的改进，分别引入了一种基于样例学习的模型匹配方法以匹配序列图像中的首幅人脸图像，并在跟踪过程中加入基于灰度归一化的光照抑制过程。实验表明改进后的跟踪过程较原跟踪过程更加准确和方便。2．在基于Candide3人脸表情跟踪过程中，提取模型参数并集合成为动态特征。首先本文中提出了一种基于六模型参数的动态特征提取方法。后通过分析表情变化时所牵动的表情运动单元变化。在原动态特征提取方法的基础上又进一步提出了一种七模型参数的动态特征提取方法，并分别对两种方法进行了聚类分析，分析结果表明基于模型参数的动态特征提取方法就有良好的聚类效果，其中进一步改进后的七参数动态特征提取方法具有更好的效果。3．在完成了动态特征提取之后，还需要通过设计分类器来最终完成表情的识别工作。分类器的设计是除特征提取以外另一个决定分类准确率的重要环节。研究中设计了多个分类器，包括K-NN分类器，SVM分类器，以及应用Adaboost算法级联贝叶斯分类器、线性判别分类器、K-NN分类器、SVM分类器等弱分类器。并通过实验，客观的给出了分类的准确率。4．最后针对提出的非定长动态特征提取这一概念给出了说明，并尝试提出了一种基于相关性的图像序列分割方法，然后应用插值的方法归一化特征长度。更多还原

【Abstract】 Face recognition involves expression of artificial intelligence, pattern classificationand image processing, psychology and so on many domains. It is an important directionabout the study of human interaction. Along with the computer running speedcontinuously improved, people pay more attention on the facial expression recognitionbased on dynamic image sequenceIt is different between the facial expression recognition based on the dynamic imagesequence and the facial expression recognition based on static image. More information isinvolved in the dynamic features. This paper outlined the research contents are as follows:1．In this part, we first studied the tracking theory based on Candide3facial model andaccording to the research we further developed the tracking process. A semi-automaticfacial model initialization method is put forward here. Then based on the semi-automaticfacial model initialization method, an example-based model initialization method isutilized. Last a normalization process is added in the whole tracking process to restrain theeffect of illumination.2．Based on the research of facial tracking which I have mentioned, I put forward adynamic feature extraction method based on six parameters of facial model. Then in orderto enhance the recognition accuracy, we further studied the movement of facial action units during different expression process. According to this research a dynamic featureextraction method based on seven parameters of facial model was developed.3． After the dynamic feature extraction, another important element of facialexpression recognition—classifier design—is considered. Various classifiers are designedhere and an objective judgment of recognition accuracy is stated.4．An unfixed size dynamic feature extraction conception is put forward and explainedand according to the conception I made a preliminary study.更多还原