【作者】 杨勇；

【导师】 王国胤；

【作者基本信息】 西南交通大学 ， 计算机应用技术， 2009， 博士

【摘要】 包括情感计算和情感识别在内,以人为中心的,对人的情感和认知的研究是目前人工智能领域的一个热点研究方向。对这一方向的研究将有助于推动实现人与机器之间的自然交互。虽然目前情感计算和情感识别的研究已经取得了初步的成果,但是由于缺乏坚实的心理学和认知科学的理论基础,在研究中还存在诸多待解决的关键问题。本文就人脸表情识别系统所涉及的特征选择和识别方法进行研究。这些研究工作不仅能推动计算机表情识别、计算机情感仿真等的进展,而且对于自然人机交互、远程教育、医疗看护、游戏娱乐等应用领域都有重要意义。本文以粗糙集理论为基础,研究了粗糙集属性约简算法,并把粗糙集属性约简算法作为一种人脸表情识别系统的特征选择方法。根据人脸几何特征直观性好的优点,以人脸的几何特征作为研究对象,对人脸表情识别的重要特征进行研究。此外,本文还基于粗糙集理论,结合集成学习理论,对人脸表情识别方法进行研究。取得的主要研究成果如下：1)提出了基于粗糙集属性约简的人脸表情识别的特征选择方法。本文引入粗糙集理论,提出了采用粗糙集属性约简算法作为一种人脸表情识别的特征选择方法。基于经典粗糙集理论,采用基于信息熵的属性约简算法和特征选择的属性约简算法对人脸表情识别采用的特征进行了特征选择,并提出了一种RS+SVM的人脸表情识别方法。研究结果表明,相比较传统的特征选择算法,粗糙集属性约简方法不仅能有效地降低特征的维度、得到人脸表情识别的重要特征,而且基于这些重要特征,能得到较好的人脸表情识别结果。2)提出了自主式的人脸表情特征选择算法。本文基于粗糙集理论和面向领域的数据驱动的数据挖掘模型思想,对连续值信息系统的属性约简方法进行了研究,提出了一个自主式的属性约简算法,首先,将条件属性集对决策的可区分性作为知识的一种属性特征,并基于粗糙集理论提出一种可区分性的度量方法。然后,根据在知识获取过程中可区分性的不变性原则,提出了一个针对连续值信息系统的自主式属性约简算法(SARA),该算法可以从训练数据中自动获取控制学习过程所需要的参数。研究结果表明,该方法在不依赖于领域专家先验知识的情况下也能得到很好的结果。本文把SARA作为一种人脸表情识别的特征选择方法,把SARA挑选的表情特征应用于人脸表情识别,得到了很好的人脸表情识别结果,同时还得到了在人脸的几何特征中嘴部的特征对人脸表情识别起到最重要作用的结论。3)基于粗糙集理论和集成学习理论,提出了一种选择性集成的人脸表情识别方法。首先,基于经典粗糙集理论的可辨识矩阵求出信息系统的多个约简,并对得到的多个约简训练多个候选分类器。然后,基于双误的差异性度量方法,对候选分类器按差异度进行聚类。最后,采用选择性集成的策略,从每两类候选分类器中选择出差异性最大的分类器进行相对多数投票,得到集成的输出结果。仿真实验结果表明,该方法得到了良好的表情识别结果。4)基于扩展的粗糙集理论和集成学习理论,提出了一种动态选择的人脸表情识别方法。首先,基于面向领域的数据驱动的数据挖掘模型思想,提出了一个求核属性的算法和一个基于信息熵的求多个属性约简的算法。然后采用该方法求出信息系统的多个约简,并根据得到的多个约简训练候选分类器集合。最后,根据多分类器的本地精确性,基于动态选择的策略从候选分类器集合中挑选出最优的分类器对待识别的样本进行分类识别。该方法成功用于直接处理人脸表情特征的连续值属性,仿真实验结果表明,该方法得到了良好的表情识别结果。5)开发了一个语音和视频的双模情感识别系统基于本文的研究成果,开发了一个语音和视频的双模情感识别系统。该系统能实现实时的单模人脸表情识别、单模语音情感识别和双模情感识别。实际测试结果表明,该系统具有很好的实时性和识别结果。更多还原

【Abstract】 The study of human centered emotion and cognition, including affective computing and emotion recognition, is a hot research topic in artificial intelligence. Achievements on these research topics will push the development of human-computer intelligent interaction (HCII). Although there are already some achievements of affective computing and emotion recognition in recent years, there are still some key problems unsolved due to the absence of solid theory basis of psychology and cognition. In this thesis, feature selection methods and emotion recognition methods in a facial emotion recognition system are studied. These research works can not only push the development of emotion recognition and emotion simulation, but also are important for the applications in HCII, E-learning, medical caring, game, etc.In this thesis, attribute reduction algorithms are studied based on rough set theory and used for feature selection of a facial emotion recognition system. Geometrical face features are taken as the research objects since it is intuitionistic. Important features for emotion recognition are studied based on attribute reduction algorithms. Furthermore, efficient emotion recognition methods are also studied based on rough set theory and ensemble learning. The major achievements of this thesis are as follows.1) A feature selection method for emotion recognition based on rough set theory is proposed.In this thesis, the attribute reduction algorithms based on rough set theory are introduced and proposed for feature selection methods for emotion recognition. Based on traditional rough set theory, the attribute reduction algorithm based on conditional entropy and the attribute reduction algorithm based on feature selection are used as feature selection methods for emotion recognition. Meanwhile, a novel emotion recognition method based on RS plus SVM is proposed. The experiment results show that the attribute reduction algorithms can reduce the feature dimension and extract the important features for emotion recognition. Based on these features, better recognition result than traditional feature selection methods can be achieved.2) A self-learning expression feature selection algorithm is proposed.In this thesis, based on the idea of domain-oriented data-driven data mining theory, an attribute reduction algorithm for continuously valued information systems is studied, and a self-learning attribute reduction algorithm is proposed. Discernable-ability of conditional attribute set with respect to decision attribute set is taken as an important property of knowledge, and a measurement of discernable-ability is proposed based on rough set theory. According to the criterion that discernable-ability should hold in the course of knowledge acquisition, a self-learning attribute reduction algorithm (SARA) is proposed for continuously valued information systems. Its parameter can be obtained automatically from training data set. Experiment results show that the method can get good result even if there is no prior domain knowledge. SARA is taken as a feature selection method for emotion recognition in this thesis, and good recognition result is achieved based on the features it found. Furthermore, the features which concerned mouth are found as the most important expression features.3) A selective ensemble feature selection method for emotion recognition is proposed based on rough set theory and ensemble learning.At first, an algorithm based on discernibility matrix is adopted and multiple reductions are generated, correspondingly multiple candidate classifiers are trained. Secondly, the double fault method is taken as the measure of the diversity of the candidate classifiers, and the candidate classifiers are clustered according to measurement of the double fault method, and the most diverse classifiers are selected from each pair of clusters. At last, the criterion of relative majority voting is adopted for the selected classifiers, and the output of the ensemble is gotten. Experiment results show that good recognition result can be achieved based on the proposed method.4) A dynamic ensemble feature selection method for emotion recognition is proposed based on extended rough set theory and ensemble learning. At first, an algorithm for calculating the core of a decision table is proposed based on the domain-oriented data-driven data mining theory. Secondly, an algorithm for finding multiple reductions based on conditional entropy is proposed. Accordingly, multiple candidate classifies are generated using this algorithm. At last, a dynamical selective measure is used to select the most suitable classifier for each unseen sample according to local property of the candidate classifiers, and recognition result is gotten accordingly. Experiment results show that good recognition result can be achieved based on the proposed method.5) An audio and visual double module emotion recognition system is developed.A double module emotion recognition system based on both audio and visual information is developed. It can recognize facial expression, emotional speech and expression of the combination of facial and speech emotions in time. Good recognition results are achieved in real testing environments.更多还原