【作者】 樊引刚；

【导师】 王方石；

【作者基本信息】 北京交通大学 ， 软件工程， 2011， 硕士

【摘要】 面部表情识别是计算机科学与人机交互领域的一个新兴研究课题。面部表情是人们内心情感的真实表露。美国著名心理学家阿尔培特认为,在人们进行感情表达时,往往言词的使用只占7%,声调占38%,而剩下的55%由面部表情来完成[1]。人脸表情识别的目的是让人工智能产品能够识别出人的表情,进而分析人的内心情感。人脸表情识别对人机交互、安全、机器人、通信和汽车领域都起着不可估量的作用；因此,对于人脸表情识别的研究具有非常重要的研究和应用价值。表情识别主要分为两大类：动态视频的表情识别和静态图像的表情识别。本人针对静态图像的表情识别做了如下工作：1.研究并实现了基于模板匹配方法的人脸检测,包括制作人脸模板,选择匹配算法及采用马赛克检验方法检测人脸；在表情识别中,成功检测人脸是表情识别的首要任务。2.研究并实现了表情图像的预处理算法,由于各种图像尺寸不同,背景不同,并会有多种光照条件的影响,因此,首先采用直方图均衡化等方法对图像进行预处理。预处理为提高表情识别率奠定了良好的基础。3.实现基于轴对称性的人眼和嘴巴的定位方法,假设单眼区域和嘴巴区域均具有关于中心垂线轴对称的特性,首先找到一只单眼的下半部分区域,迭代选择候选眼睛中心,再利用轴对称性确定整只眼睛的区域,同理可确定另一只眼睛区域及嘴巴区域。在表情识别中,能否准确定位眼睛和嘴巴直接影响到特征值的提取和识别率。4.分别针对高兴、厌恶、悲伤、生气、惊奇、害怕和中性七种表情建立隐马尔可夫模型。在表情识别过程中,由于数据量大、计算复杂度高,会直接影响系统的反应速度和识别率；K-L变换可以把数据从高维空间映射到低维空间,提取出原始数据中的主要特征；因此,本文对特征区域信息进行K-L变换,提取主要特征向量,减小数据计算量及计算复杂度；并将特征向量输入到隐马尔可夫模型中,通过多次跌倒叠代,训练出相应的、稳定的表情模型；在识别过程中通过计算给定图像与训练好的模型的相似度,确定相似度高的模型所代表的表情即为给定图像的表情。系统在测试过程阶段,表现出良好的性能：以JAFFE和索尼(中国)研究院人脸表情数据库为实验数据集,采用模板匹配和马赛克检验方法进行人脸检测,其检测正确率可达到96%；采用K-L变换预处理图像,降低了眼、嘴特征向量的维度,其降维比例可达到65536：3或者65536：6,且信息损失的均方差不超过15%；采用隐马尔可夫模型进行表情识别,针对同一对象进行表情识别,其准确率可达87.55%,针对不同对象进行表情识别,其准确率可达到76.95%。更多还原

【Abstract】 Facial expression analysis is rapidly becoming an area of intense interest in computer science and human-computer interaction design communities. The most expressive way humans display emotions is through facial expressions. One of America famous psychologist named Albert said:when a man shows his emotions, his words carrying 7 percent of emotions, his tune carrying 38 percent and his expression carrying the rest 55 percent. The purpose of facial expression recognization is to make the computer recognize people’s facial expressions based on which their emotion status will inside then be analyzed and identified. Facial expression recognization has wide range of applications in the man-machine interaction, robot making, security, medical treatment, communication and automobile areas, etc. Therefore, it is very important to do the facial expression research for its application value in the future.The method of facial expression recognization contains two categories:facial expression recognization in video and static picture. The paper contains following research and implementation on static picture expression recognization:First, the algorithm is designed and implemented based on template matching, which contains template maked, matching method choosing and mosaic testing. To detect people’s face successfully is the first step in an automatic facial expression recognization system.Second, some pretreatment methods for expression recognization are researched and implemented. In different pictures, there are various poses, size, background and light conditions. Histogram equilibria, pictures unification and many other pretreatment methods are used to pretreat the pictures in this paper, which builds a good base for the facial expression recognization.Third, an eye and mouth location method based on the axial symmetry is researched and implemented. Assuming monocular area and mouth area have the vertical axisymmetric characteristics, the lower area of one eye is found firstly and the candidate eye center will be iteratively selected, then the whole eye area will be determined by axisymmetric property on eye area; the other has also the same method. An important step in the facial expression recognization is the eyes and mouth locating.Finally, the paper applies the HMM (Hidden Markov Model) in expression recognition, and makes each model for each expression (happy, surprise, sad, anger, disgust, fear, neutral).There are massive data needs to be handled during the expression recognition, and the calculation process is quite complex, which will impact the performance and the recognition rate of the system. Therefore the paper uses the K-L method to transform the eyes and mouth area, extracts the main features, Reduce data calculation and complexity; and then inputs the features to the HMM, trains many times and builds robust models. In the recognization process, it calculates the similarity between given picture and models, and then takes the high similarity model’s expression as the final result.The System showed good performance in the testing. In the testing of template matching and mosaic, the detection rate is up to 96% in the facial expression database of the JAFFE (Japan) and Sony (China) Institute. In the testing of feature location and extraction, the use of K-L transform made original data dimension reduced (65536:3 or 65536:6) and the mean square error rate was less than 15% after the transform. In the testing of expression recognition, this paper applies the HMM (Hidden Markov Model) in the expression recognition, the algorithm achieves an accuracy of 87.55% in the same subjects and 76.95% in the different subjects for facial expression recognition from the facial expression database of the JAFFE (Japan) and Sony (China) Institute.更多还原