【作者】 周书仁；

【导师】 梁昔明；

【作者基本信息】 中南大学 ， 计算机应用技术， 2009， 博士

【摘要】 人脸表情识别在医学和民用领域有着广阔的应用前景,是目前的一个非常活跃的研究领域。人脸面部表情蕴涵着丰富而又细腻的情感和心理信息,面部表情识别主要涉及两个方面的问题,如何有效地获取人脸面部表情特征和如何开展表情分类分析研究。本论文主要研究内容与创新性工作包括以下五个方面:1)全面综述了人脸表情识别研究的历史和现状,阐述了人脸表情识别的三个环节相关问题,即表情识别的特征提取、分类和识别框架。在表情特征提取环节,总结了一些常见的表情特征提取方法;描述了表情图像预处理中常用的一些方法;分析了Gabor与Adaboost方法的特征提取基本思路;论述了主动形状模型与主动外观模型进行表情特征提取的算法描述;验证了采用基于流形算法进行脸部特征提取的可实现性。在表情分类环节,讨论了传统的两分类支持向量机可用于表情多分类的解决过程;给出了表情数据可以通过聚类分析算法进行先聚类后区分的思想,以达到最终的表情分类识别;引用了自相似传播聚类的算法思想,阐述其解决寻找最优表情分类中心的一般过程。在表情识别框架环节,设计了表情识别系统的一般流程框架。2)提出了基于Gabor特征与类间学习神经网络的表情识别算法。首先对Gabor小波特征与BP神经网络相结合的表情识别方法进行了研究与分析,并讨论了BP网络结构设计的一般模型,分析了隐层节点数选择的方案、网络常用传递函数和如何在实际问题中加以应用,同时对神经网络模型的学习过程进行了详细阐述,为类间神经网络的具体提出做好铺垫;然后提出一种类间学习神经网络表情识别算法,该算法给出了如何获取神经网络输入层所需要的特征,如采用Gabor小波变换提取人脸表情的局部特征,且只选择对表情贡献最大的特征区域,建立了类间学习神经网络的一般概念,阐述了类间学习神经网络的结构描述,并设计了类间学习神经网络的模型,分析了类间学习神经网络的训练过程,描述了表情识别实现思路;其次,针对局部Gabor特征与神经网络的表情识别方法进行了相关的实验,通过实验给出了一般表情对之间的距离判距,验证了通过类内期望输出和类间期望距离来修正网络的正确性;最后给出了一组实例样本进行表情分类识别。3)提出了基于ICA特征与隐马尔可夫模型的表情识别算法。首先介绍了基本的ICA模型,根据表情图像的数据输入空间构造ICA模型,描述了FastICA算法步骤,分析了FastICA算法求解分离矩阵的过程,采用了滑动子窗口提取训练集特征,得到了人脸表情基向量;然后介绍了粒子群优化算法步骤,利用其在数据空间搜索最优解的优良表现,将其引入到表情特征的快速求解上,避免了求解过程中的复杂计算开销,并比较分析了优化后的ICA与基本ICA进行特征提取所具有的不同表现;其次,介绍了HMM的一般概念,给出了HMM表情识别建模的详细过程,在确定HMM模型的各个参数上进行了实验分析比较,以落实最佳的参数为识别分类所用,给出了HMM的一般训练过程和基于优化后的ICA与HMM算法的表情识别体系;最后,比较分析了相近方法的研究,得出一些影响结果变化的因素,通过实验结果验证本文所提出思路的正确性。4)提出了基于活动外观模型特征与Adaboost的表情识别算法。首先介绍了ASM方法和AAM方法的一般概念;然后针对表情图像的标记点和轮廓线的构成给出了依据,并比较了ASM和AAM的不同,阐述了利用AAM提取人脸表情特征所具有的优势;其次,给出了Adaboost多分类算法步骤描述,为了能够将AAM获取的特征点值用于Adaboost多分类,构造了一种适合于Adaboost多分类所需要的Harr型特征,并应用此特征通过Adaboost分类器进行表情识别;最后,通过实验对训练集错误率等数据进行了分析,并针对不同方法之间的性能作出了分析比较。5)提出了基于流形特征与支持向量聚类的表情识别算法。首先针对目前用于表情识别主要的两类流形学习算法进行了介绍,根据LPP降维算法提出了一种有效的有约束机制,并在实验中验证了这种约束能够更好地优化投影矩阵的选择;然后给出了一种LCSVC的聚类过程,有效降低了表情类别的聚类边缘的部分干扰,在SVC聚类过程,采用MFA方法调整各个SV的权重;其次,设计了一个四层神经网络表情分类器,这一设计思路是一种表情识别方法的有效途径;最后,验证了系统中一些重要环节用到的参数,分析了不同方法之间的性能比较。更多还原

【Abstract】 Facial expression recognition has broad prospects in the field of medicine and civilian applications, nowadays it becomes a hot research field. Facial expressions implicate abundant and exquisite emotional and psychological information. Facial expression recognition is mainly involved with two issues, which are how to obtain the features of facial expression and how to research expression classification.In this dissertation, the main research contents and innovative work include the following five aspects:1) The dissertation provides a comprehensive overview about the history and current situation of facial expression recognition research. It also illustrates three procedures in this field, which are features extraction, facial expression classification and expression recognition framework. Firstly, in the procedure of expression features extraction, the dissertation summarizes some methods of general expression features extraction, and describes some common methods of expression images pretreatment. It analyzes the idea of features extraction based on Gabor wavelet transform, discusses the method of features extraction based on active shape model and active appearance model, and validates the realization of features extraction based on manifold algorithm in this procedure. Secondly, in the procedure of facial expression classification, the dissertation discusses the traditional method based on two classification support vector machine, and illuminates an idea in detail that the expression data can be classified after clustering using cluster algorithm, and then realizes facial expression recognition. It analyzes the algorithm of affinity propagation, and describes the general process of the optimized expression classification center. Thirdly, the general flow framework of expression recognition system is designed in the procedure of expression recognition.2) The dissertation proposes a facial expression recognition algorithm based on Gabor features and congener learning neural network. Firstly, the dissertation researches and analyzes the method of Gabor wavelet features combined with BP neural network, discusses the general model of the BP network structure, and analyzes the project of the decision of hidden layer nodes number, the network transfer function and their application in practice, meanwhile it expatiates the learning process of neural network model aiming at paving the way for congener learning neural network. Secondly, an algorithm of congener learning neural network is proposed in the dissertation, which provides a method to obtain the required features of the neural network input layer, for example, Gabor wavelet is adopted to extract local facial expression features, and the feature region which contributes the most to the facial expression are only selected, then the general concept of congener learning neural network is built. The dissertation illustrates and designs the structure and model of congener learning neural network, it also analyzes the training process of congener learning neural network and describes the implementation of expression recognition. Thirdly, according to the method of local Gabor features and neural network facial expression recognition, some related experiments are tested. The results of experiments provide the distance judgment evidences between two general expressions, and prove the correctness of the method which amends the parameters through the output of inner class expectation and congener expectation distance. In addition, a group of samples are tested for expression classification.3) The dissertation proposes a facial expression recognition algorithm based on ICA features and HMM. Firstly, it introduces the basic ICA model, builds ICA model according to the expression image data, describes the FastICA algorithm procedure, analyzes the process of FastICA algorithm solving separation matrix, utilizes the sliding sub-window to extract training set features, then gets the facial expression basis vector. Secondly, it introduces the PSO algorithm which has outstanding performance of searching optimal solution in the data space. This method avoids the complex computation in extracting expression features. The dissertation shows that the optimized ICA and basic ICA for features extraction have different manifestations. Thirdly, it introduces the general concept of the HMM and the detailed process of HMM expression recognition model, and then it compares and analyzes experiments results based on various parameters of HMM model in order to get the best parameters for classification. It provides the general training process of HMM and the expression recognition system based on the optimized ICA and HMM algorithm. Finally, it compares and analyzes some similar methods and draws a conclusion that some factors will affect the experimental results. Through experimental results, the dissertation verifies that the proposed ideas are correct.4) The dissertation proposes a facial expression recognition algorithm based on AAM features and Adaboost expression classification. Firstly, the ASM method and its general concepts are introduced. Secondly, it provides the evidence according to the marker of expression image and contour, compares the differences between ASM and AAM, and then illustrates the advantages of using AAM to extract facial expression features. Thirdly, it explains Adaboost multi-classification algorithm procedure. In order to apply the features point value of AAM in Adaboost multi-classification, the dissertation constructs the Harr-like features which are quite suitable for Adaboost multi-classification, and through Adaboost classifier it can recognize the facial expression. Finally, through experiments the error rates of training set are analyzed, and the performances are compared in different methods.5) The dissertation proposes a facial expression recognition algorithm based on manifold features and support vector clustering. Firstly, two manifold learning algorithms are introduced, and then an effective and restricted mechanism is proposed according to LPP descending dimension algorithm, meanwhile experiments validate this mechanism can optimize the choice of projection matrix. Secondly, a process of LCSVC clustering is provided which can reduce the interference in clustering marginal of expression classification. In the process of SVC, the method of MFA is adopted to adjust the weight of each SV. Thirdly, a facial expression classifier with four-layer network is designed which is an effective approach to classify expression. At the end, the dissertation validates the parameters in some important procedures and compares the performances in different methods.更多还原