【作者】 李亚东；

【导师】 阮秋琦；

【作者基本信息】 北京交通大学 ， 模式识别与智能系统， 2010， 硕士

【摘要】 人脸表情含有丰富的人体行为信息,在人类非语言方式的交流中起到了主导作用,近几十年来在人机交互领域也受到越来越多的关注。若能使计算机拥有更强的识别和理解人脸表情的能力,将会极大的改变计算机与人的关系,从而使计算机更好的为人类服务。本文对国内外关于人脸表情识别的文献进行了深入研究和分析,针对表情识别的若干问题进行了探讨,并且对在识别过程中占据着重要地位的特征提取的各种算法做了深入的研究,在此基础上提出了几种改进的算法。大量的实验证明本文提出的算法的高效性。主要工作如下：第一,改进了现有的有监督的谱特征分析算法。在原有算法的基础上,引入了另外两种Laplacian矩阵,并通过实验验证了算法的有效性。第二,提出了基于双元树复小波变换的有监督的谱特征分析算法。引入双元树复小波变换并进行了改进,利用其平移不变性、方向选择性、完全重构性和高效计算能力等特性,对图像进行4层分解来提取表情特征。每层分解得到6幅指向不同方向的带通子图,体现了其多尺度多方向的分辨能力,使得表情中细微的局部特征更好的体现出来,并使得识别率有了大幅提高。然后结合有监督的谱特征分析算法进行表情识别,在JAFFE库和CK库上通过大量的实验来验证本文算法的有效性。第三,引入单元树复小波变换与双元树复小波变换进行比较,同时提出基于单元树复小波的有监督的谱特征分析算法。对比得知两种复小波变换虽然表征图像特征的能力很相似,但单元树复小波变换计算相对复杂。第四,搭建了人脸表情识别演示系统,将本文研究的算法的识别结果更直观的显示出来,同时为后续研究提供了一个实验平台。更多还原

【Abstract】 Facial expression holds rich human behavior information and plays a leading part in human non-verbal communication. In the field of Human-Computer Interaction, more attentions have been paid on it in recent decades. If computer is able to recognize and understand human emotions better, human and computer interaction will be greatly improved and serves human better.In this paper, the literatures home and abroad about facial expression recognition are in deep research and analysis, and it shows that feature extraction gets much importance in recognition system. We focus a number of issues about facial expression recognition and make research on facial expression feature extraction algorithms. Furthermore, some improvements of feature extraction methods are proposed. Experiments prove the efficiency of the proposed algorithms. The main contributions of this paper are:Firstly, we improve the existing method of supervised spectral analysis. Another two Laplacian matrixes are introduced to get better performance, and experiments are done to verify its effectiveness.Secondly, supervised spectral analysis based on dual-tree complex wavelet transform is proposed. We introduce the dual-tree complex wavelet transform and bring forward some improvement. In virtue of the attractive properties such as shift-invariance, direction selectivity, perfect reconstruction and efficient computing, we decompose the image to 4 scales to extract expression feature.6 sub-images oriented at 6 different directions will be obtained at each scale, and it shows its multi-scale and multi-direction resolution ability. Meanwhile subtle and local characteristics of image will be presented better. Combined with supervised spectral analysis, the dual-tree complex wavelet transform enhances the recognition performance remarkably. Experiments on the JAFFE database and CK database show its effectiveness.Thirdly, the dual-tree complex wavelet transform and single-tree complex wavelet transform are compared with each other. Furthermore, we propose the method of supervised spectral analysis based on the dual-tree complex wavelet transform. By comparison of these two complex wavelet transforms, we learned that though the ability to present image features is very similar, the single-tree complex wavelet transform is more complex to compute.Fourthly, a demonstration system for facial expression recognition is built to show our achievement, and an experimental platform is provided for the follow-up study at the same time.更多还原