【作者】 吕兴会；

【导师】 郑秋梅；

【作者基本信息】 中国石油大学 ， 计算机科学与技术， 2010， 硕士

【摘要】 人脸表情识别是人工智能与人机交互领域中的一个重要课题,目的是实现计算机自动识别人的表情,进而分析人的情感与心理。这将进一步增强人机交互的友好性与智能性,因此有着很高的研究价值和良好的应用前景。人脸表情识别系统一般包括人脸检测、特征提取、表情分类等环节。本文主要研究了特征提取和表情分类过程中的一些关键问题,分析了现有算法中的不足,提出改进算法,并进行了仿真实验。主要的工作如下:第一,改进Gabor特征提取算法。提出一种Gabor特征选择方法,利用分布估计算法对Gabor滤波器组的尺度和方向进行优化选择,选择出最佳的Gabor核的尺度和方向,并在此基础上提取特征。在日本的JAFFE表情数据库上的实验结果表明,无论在计算量还是识别性能上都比传统的Gabor滤波器组更具有优势。第二,提出双向二维直接线性判别分析算法。该算法从水平和垂直两个方向对图像矩阵执行直接线性判别分析,进行列和行的两次维数压缩,降低了特征维数。在JAFFE人脸表情数据库中的实验结果表明,算法在提高了降维能力的同时,也提高了识别率。第三,将二维判别保局投影算法应用到人脸表情识别的特征提取中。在JAFFE表情库的实验中,分析比较了二维判别保局投影算法较之于二维保局投影算法在表情特征的识别与分类中的异同及优势。第四,提出一种基于多特征集成分类器的方法进行人脸表情识别。构造一个基于神经网络的集成分类器模型,对多特征多分类器的输出进行决策融合。在JAFFE表情库的实验中取得了令人满意的结果,体现了多特征集成分类器的明显优势,也充分说明了神经网络的集成分类器模型的准确性和稳定性。实验结果表明,本文改进的Gabor特征提取算法、提出的双向二维直接线性判别分析算法,在提高了降维能力的同时,也提高了识别率。提出的多特征集成分类器的方法在JAFFE表情库可以实现91.43%的最高识别率,识别效果令人满意。更多还原

【Abstract】 Facial expression recognition is one of the most challenging research topics in the fields of artificial intelligence and human-machine interaction. Aiming at letting computers recognize human facial expressions automatically. Consequently analyzes emotions and psychology. This will further strengthen the friendliness and intelligence of human-computer interaction. It has both high research value and wide range of potential market value.Facial expression recognition system consists of such modules as face detection, feature extraction and expression classification. This paper mainly studies a number of key issues in the process of feature extraction and expression classification, analyzes the deficiencies in the existing algorithms. Several improved algorithms and methods for these tasks are developed. And simulation experiments are did. The major contributions of this paper are as follows:Firstly, improve the Gabor feature extraction algorithm. We propose a new Gabor features dimension reduction method that utilizes estimation of distribution algorithms (EDA) to search optimal Gabor kernels’scales and orientations. Experimental results on JAFFE database demonstrate that our method is more effective for both dimension reduction and image representation than traditional Gabor filter bank.Secondly, propose two directional two dimensional direct LDA((2D)2DLDA) algorithm. DLDA algorithm first works in row direction of image and then works in the column direction of image to directly extract the image scatter matrix from 2D image. Compress the dimensions from column direction to row direction, reduce the feature dimensions. Experimental results on JAFFE database demonstrate that our proposed algorithm not only enhance the ability to reduce dimensions, but also improve the recognition rates.Thirdly, apply the two-dimensional discriminant locality preserving projections ,(2D-DLPP) algorithm. Experimental results on JAFFE database show that 2D-DLPP is better than 2DLPP and LPP in expression feature extraction and classification. Lastly, propose a multi-feature and combining multiple classifiers method for facial expression recognition. We develop a model of combining multiple classifiers based on nerve net. The outputs of three classifiers are input to the model to get facial expression recognition. Experimental results on JAFFE database demonstrate that our proposed method is superior to the single feature and single classifier.The experimental results show that our improved Gabor feature extraction algorithm and proposed two directional two dimensional direct LDA algorithms not only enhance the ability to reduce dimensions, but also improve the recognition rates. Our proposed multi-feature and combining multiple classifiers method get satisfactory results. The high recognition rate reaches 1.43% on JAFFE database.更多还原