【作者】 朱明旱；

【导师】 罗大庸；

【作者基本信息】 中南大学 ， 控制科学与工程， 2009， 博士

【摘要】 表情是人类用来表达情绪的一种基本方式,是非语言交流中的一种有效手段。人们可以通过表情准确而微妙地表达自己的思想感情,也可以根据表情来辨认对方的态度和内心世界。随着人脸检测、人脸跟踪和人脸识别技术的不断完善,针对人脸表情识别技术的研究,已成了模式识别与人工智能领域的研究热点。目前,这项技术仍处在研究阶段,许多问题还需要进一步的研究才能解决。例如：如何准确地提取面部的表情特征,如何降低身份特征对表情特征的干扰,如何较好地实现表情特征的分类,如何估计表情强度,以及如何对混合表情进行分析等等。由于表情变化具有非线性和连续性的特点,本文采用了流形学习算法来对人脸表情的变化特征进行提取。主要工作集中在两个方面：一是人脸表情类别识别的研究,主要是通过表情的特征提取和对提取特征的分类,来实现表情图像的类别判断。根据样本标注信息的不同,本文分别研究了无身份标注和有身份标注下的表情类别识别。二是人脸表情成分分析的研究,主要是分析表情图像中,包含的各种基本表情成分的强度。本文的创新点和主要贡献如下：1)在无身份标注的表情识别中,提出了一种新的监督式等距映射算法。该算法在实现表情原始特征降维的同时,较好地实现了相同对象的表情图像和相同类别表情图像的双重聚类。减少了表情分类时,身份特征对表情特征的干扰,以及不同表情图像间的干扰。2)在无身份标注的表情识别中,针对表情流形特征的分类,提出了一种序列加权k近邻分类方法。该方法在对某待测样本进行分类时,有效地利用了表情流形结构的特点,避免了用加权k近邻分类器时出现的许多错分现象,提高了分类的正确率。3)在有身份标注的表情识别中,提出了先识别身份再在同身份图像内部进行表情识别的处理方法。针对表情图像的身份识别,提出了一种广义的主成分分析算法,它是二维主成分分析和模块二维主成分分析人脸识别算法的进一步推广,它的特征提取过程不再受到图像矩阵维数的任何制约。4)针对局部线性嵌入的批处理问题,提出了一种正交迭代局部线性嵌入算法。该算法能够不断地利用前面的流形学习结果,得到新样本的流形嵌入向量,实现了局部线性嵌入的增量处理。在有身份标注的表情识别中,该算法的应用为待测图像的流形嵌入提供了极大的方便。5)在有身份标注的表情识别中,针对同身份图像内部的表情识别,提出了一种通过表情图像的重建来实现表情识别的算法。该算法通过比较待测图像和它重建图像的相似程度来实现表情分类,由于重建图像的表情强度会跟随待测图像的表情强度而自动变化,该表情算法对表情强度的变化表现出了较好的鲁棒性。6)实际生活中的表情大多是一种混合表情,因此仅将这种表情归为几种基本表情类别中的某一类,往往难以满足情感分析的客观要求。针对表情的情感分析,提出了一种基于流形空间模型的表情成分分析方法。该方法可以将任意的表情分解为一些基本表情的矢量和,这为表情的混合成分的分析提供了一种新的处理思路。更多还原

【Abstract】 Facial expression is a basic manner to express human emotion and a powerful way of non-language communication. People can express their thoughts and feelings delicately through expression, and catch others’ attitude and inner world by virtue of expression. With the development of face detection, face tracking and face recognition technique, the study of facial expression recognition has become an attractive issue in the areas of pattern recognition and artificial intelligence. At present, the expression recognition technique is still at research phase. Many problems still need to be investigated:how to extract facial expression features exactly, how to reduce the influence of identity features on expression features, how to classify expression vectors accurately, how to define the expression intensity, how to estimate the intensity of expression, and how to analysis mixture expression etc.In view of the non-linearity and continuity of expression variation, some manifold learning algorithms are used to extract facial expression features in this dissertation. The work in this dissertation aims at two issues:one is facial expression recognition; the other is facial expression analysis. The main task of the former is to classify expression images. The aim of the latter is to analyze the intensity of each basic expression component in mixture expression. According to the label information of train samples, two kinds of expression recognition are investigated in this dissertation. One is expression recognition without identity label information; the other is expression recognition with identity label information. The main innovations are as follows:1) A new supervised Iosmap algorithm is proposed in the expression recognition with no identity label. The algorithm reduces the dimension of expression feature vectors and implements the image clustering according to their identity and expression class. This method has cut down the interference between identity features and other expression features.2) For better implementing features classification in expression recognition with no identity label, a sequential weighted k-nearest neighbor classification method is proposed. The method makes good use of the structure of expression manifold for test image classification. It avoids many misclassification situations using weighted k-nearest neighbor classifer and improves the accurate rate of classification.3) A facial expression recognition idea based on face recognition is proposed. First, the face identity is recognized. Then the facial expression is recognized among images with the same identity. A generalized principle component analysis method is proposed countering the identity recognition of ecpression images. The method is an extended algorithm of two-dimensional principal component analysis and modular two-dimensional principal component analysis. The features extraction oprocess will not be restricted by the size of image matrix.4) Aimed at the defection that locally linear embedding is a batch processing algorithm, a locally linear embedding algorithm based on orthogonal iteration is proposed. The method can use the former results of manifold learning continually to compute the manifold vector of new sample. It is an increment manifold embedding algorithm. In the expression recognition with identity label, the method affords many facilities for test images manifold embedding.5) An expression recognition algorithm based on image reconstruction is proposed aimed at a specified subject’s expression recognition in expression recognition with the identity label. The method classifies test image according to the similarity between it and its reconstruction images. As expression intensity of reconstruction images will change with test images, the algorithm is robust to expression intensity.6) Many facial expressions in daily life are mixture expression. It does not satisfy the demand of emotion analysis by classifying mixture expression into just one special class of the basic expressions. An expression component analysis method based on manifold space mode is proposed aiming at this demand. With the method, any mixture expression image can be resolved into a vector addition of some of basic expressions in expression manifold space. The method is a novel way to analysis mixture components of facial expression.更多还原