【作者】 李俊华；

【导师】 彭力；

【作者基本信息】 江南大学 ， 控制理论与控制工程， 2008， 硕士

【摘要】 在过去的二十年里,人脸表情识别是模式识别和人工智能领域的研究热点,在人机互交中起着重要的作用。计算机和机器人如果能够像人类那样具有理解和表达情感的能力,并能够自主适应环境,这将从根本上改变人与计算机之间的关系,使计算机能够更好的为人类服务。这也正是研究人脸表情识别并赋予计算机具有情感理解和情感表达课题的意义。许多应用都要求人脸表情识别有足够的准确度,从而完成预定的目标。例如情感分析、虚拟现实、视频会议、医学护理、商场和酒店的顾客满意度测试等等。因此,通过分析人脸表情所提供的信息,研究出鲁棒、准确、快速且智能的自动人脸表情识别方法是非常重要的。这也将推动人脸表情识别的应用越来越广。本文首先阐述了面部表情识别的课题背景及其起源、发展与研究现状,并结合国内外相关领域的发展,从特征提取与表情分类两个角度对目前表情识别的主流方法做了详细介绍。虽然人们互相之间辨别对方表情不是件困难的事情,但对于计算机而言却并非那么得简单。因为人脸表情易受年龄、种族、性别、光照等因素的影响,并且人脸是一个塑性变形体,其上的表情变化幅度也千差万别,很难建立普遍适用的人脸表情模型。特征提取在表情识别过程中非常关键,对接下来的表情分类产生至关重要的影响。根据主成分分析法,得到了适合表情识别的特征块主成分分析法,进一步把它推广到二维,形成特征块二维主成分分析法。此方法在降维的同时有效的保留了表情信息。此外,还对特征空间的维数确定问题进行了探索,得到了一些有益的结论。在特征提取方面,还提出了特征差值矩阵,这进一步消除了不同表情之间的相关性,提高了不同表情的可分性。在表情识别方面,基于Manhattan距离得到了表情分类方法——Manhattan距离分类器,与传统的欧氏距离、余弦距离分类器进行了比较,得到了比后两者更快的计算速度和更好的识别率,显示出了Manhattan距离分类器在表情识别中的优势。为了克服不同表情样本之间存在模糊性、数据交叉等情况引起识别率降低的问题,应用了量子神经网络来对表情进行分类。量子神经网络是将量子神经元与模糊理论相结合的模糊神经系统,它能将决策的不确定性数据合理地分配到各模式中,从而减少模式识别的不确定度,提高模式识别的准确性。为了进一步提高识别率,基于多分类和多层级的思想设计出了分类器组合和串级识别的方法。这时,分类错误所带来的风险可能还比较大,为了降低分类错误的风险,引入了拒识的方法,并且提出了阈值自适应调整机制根据不同的表情自动的调整阈值。在全文的结论部分,归纳了本文所做的主要工作和仍然存在的几点不足,并对今后的基于图像处理的面部表情识别研究和发展趋势做了展望。更多还原

【Abstract】 During the past two decades, facial expression recognition has attracted a significant interest in the pattern recognition and artificial intelligence due to its importance for human interfaces. Facial expression recognition plays a vital role in human-computer interaction (HCI). If computer or robot own the understanding and emotional expressing capabilities like human beings, it will thoroughly change the relationship between humankind and computer. And then, computer will provide much better services to human beings. That is also the reasons why we study facial expression recognition and endue computer with emotional understanding and emotional expressing. Many applications, such as emotion analysis, virtual reality, video-conferencing, medical nursing, and customer satisfaction studies for shop and restaurant services and so on, require efficient human expression recognition in order to achieve the desired results. Therefore, it is necessary to develop robust, accurate and fast intellectual automatic facial expression recognition system by analyzing the information of facial expressions. This will also constantly promote applied growth of facial expression recognition.To begin with, the research background of facial expression recognition is introduced, such as the research’s origin and developments. What’s more, based on the actuality of this research area, we describe a survey of facial expression recognition’s methods in terms of feature extraction and expression classification. People distinguish expression easily each other. But, it is a very defiant task for computer because facial expression has relation to age, race, sex, illumination and so forth. And face is a gentle body, rather than just rigid body, and is hard to build up an ideal and precision facial expression model to describe. Feature extraction is an important step of the whole facial expression recognition. The accuracy of feature extraction has an important affection for classification decision. In this paper, we propose feature block principal component analysis (FB-PCA) which fit expression recognition much more by studying the algorithm of principal component analysis. Furthermore, we extend this method to two dimensions and present feature block two dimensional principal component analysis (FB-2DPCA). The main expression information is reserved when the dimension is reduced. We also discuss how to determine the dimension of feature space and draw some conclusions. Otherwise, we propose a method of feature difference matrix (FDM) in order to eliminate the correlation of different expressions. In this way, different expressions can be classified more easily.In the aspect of expression recognition, we propose a new method of facial expression classification, namely, Manhattan distance classifier. Compared with the traditional Euclidean distance classifier and Cosine distance classifier, the method is proposed by us has a great improvement in calculation speed and has some improvement in recognition rate. In order to overcome the problem of indetermination and crossed data lead to low recognition rate, we propose quantum neural network (QNN) classifier. QNN is a kind of neuro-fuzzy system by merging neural modeling with fuzzy-theoretic concepts. That makes network have the fuzzy character and assign indeterminate data to all related categories. Therefore, the QNN decrease indeterminacy of pattern recognition and increase veracity of pattern recognition. Furthermore, we also propose methods of classifier combination and cascade recognition based on the concept of multi-hierarchial to improve recognition rate. But, the wrong classification still gives rise to big risk. In order to solve this problem, a method of self-adaptive threshold regulation is proposed. So threshold can be adjusted by the regulation function.Finally, we summarize our achievements and limitations, and then make an expectation for future research of facial expression recognition at the last part of the thesis.更多还原