【作者】 王国江；

【导师】 王志良；

【作者基本信息】 北京科技大学 ， 控制理论与控制工程， 2007， 博士

【摘要】 情感计算与人工心理是和谐人机交互与人工智能领域中日益受到关注的新的研究方向,也是数学、信息科学、智能科学、神经科学、以及生理、心理科学等多学科交叉的新结合点。情感计算的研究就是试图创建一种能感知、识别和理解人的情感,并能针对人的情感做出智能、灵敏、友好反应的计算系统。本文以人工心理理论和情感计算为导向,对人机情感交互的技术和方法进行了探索性的研究,探讨了人机情感交互的相关技术如情感建模、表情识别和表情合成问题,为建立和谐的人机提供相关的技术支持。本文的主要研究内容和创新点如下:(1)提出了一种结合Zernike矩与基于MCE准则的HMM的面部表情识别方法。在特征提取问题上,采用了基于人脸局部信息的矩特征提取方式,通过分割人脸局部区域,提取了最能反映表情信息的眼睛和嘴巴子图像,对子图像计算矩特征向量。在分类问题上,采用了基于MCE训练准则的HMM进行分类,实验表明,该准则函数相对于原来准则函数定义更为合理,能有效利用训练样本集中的鉴别信息,使得训练数据得到充分利用,提高了HMM的性能。(2)情感的计算模型被认为是通过与用户的情绪发生作用来实现更有效的人机交互的一个关键组成部分。在建立情感计算模型的过程中,提出了一种综合了维度论和基本情绪理论的情感数值化分类描述方法,建立了一个情感的数值化空间。并且提出了个性、心境、需要等情感成分的数值化描述方法和需要的动态处理方法,建立了各情感成分之间的映射关系。如何实现各种情感成分之间的有效组织和协调,以及实现由外部环境、情绪状态到情感行为的映射是情感建模中的难点问题。为解决这一问题,提出了基于情感认知评价理论的情感计算模型。模型采用马尔可夫过程来描述情感状态的变化转移过程,并利用马尔可夫决策过程建立了情感各成分与情感表现之间的联系。(3)在人脸表情合成部分,以Candide模型作为参数化面部模型,对网格进行纹理贴图后,通过控制AU(Action Units)来控制人脸模型的面部表情,达到参数化控制的目的,进而可以精确合成多种表情。并提出了一种自动和交互方式结合的模型调整算法。在自动提取效果欠佳的情况下可以交互修正正面特征并估计特征点深度,利用径向基函数内插来调整模型非特征点。与其他的方法相比,在图像质量好的情况下,该算法实现了自适应的调整,具有一定的实用价值。在上述建立的各种算法和模型基础上,设计和实现了拟人化网络型信息咨询服务系统,验证了模型的有效性和正确性。文中的情感模型使得信息咨询服务系统能够表现出丰富的情感。和一般的信息咨询服务系统相比,该信息咨询服务系统更具智能,更有拟人化的特征。更多还原

【Abstract】 Affective computation is a new field that has received growing interest in harmonious human-computer interactive and artificial intelligence. It also is a new union spot of multi-disciplinary such as mathematics, information science, intelligent science, nerve science, physiology and psychology etc. The research of affective computation is attempts to found a computational system which has capable of perception, recognition and understand human’s emotion and makes intelligently, keenly, friendly response aiming at huamn’s emotion. Based on affective computation and artificial psychology, this thesis discusses in detail facial expression, affective modeling and expression animation and affords technique supports for harmonious human-computer interaction. The main contributions of this thsis include:(1) A new facial expression recognition method based on zernike moment and the minimum classification error (MCE) trained HMMs is presented. In the feature extraction problem, the method of moment invariants feature extraction based on local feature regions is adopted. First, eyes and mouths are segmented from the facial expression image and variant moments feature vectors of eyes and mouth are extracted. In the classification problem, by adopting a more realistic MCE definition, discriminative information contained in the training data could be used to improve the performance of HMM.(2) Emotion model is a key part for realizing a more effective human-computer interaction.in the prossess of building emotion computational model, a classification method which synthesizes dimension theory and basic emotion theory is presented and an emotion space is built. This thesis presents numerical description of some emotion ingredient such as personality, mood, and need and dynamic processing method of need. Moreover, the mapping relationship among various emotion ingredients is built. How to realize effective organization and coordination among each kind of emotion ingredient, as well as How to realize mapping from external environment and emotion states to emotion behavior are difficult question in the emotion modelling. In order to solve this problem, emotion computation model based on the emotion cognition appraisal theory is proposed. The model describes the shift of emotion states by Markov process, and builds the relation between emotion ingredients and emotion expression by the Markov decision processes.(3) In the expression synthesis, Candide model is adopted as facial model and Face animation code system is adopted to divide the of the face into 12 AUs (Action unite) which can synthesize most expressions. the thesis present a synthesis method of model adjusted that combining automation and interaction mode. First eigenpoints are extracted by deformable templates, which can be replaced by manual manipulation when the image’ s quality is bad. Second the model is adapted globally and locally and the noneigenpoints is interpolated by RBF.In the thesis, an anthropopathic information consultation service system is built based on the models and approachs presented above and the correctness and validity of these models and approachs are proved by the system. Compared with other information consultation service system, the system is more intelligent and has anthropopathic characteristic.更多还原