基于序列特征的随机森林表情识别

发布时间：2018-04-17 10:30

本文选题：表情识别 + 随机森林　；参考：《电子科技大学》2013年硕士论文

【摘要】：人脸表情是人际沟通过程中一种非常重要的信息表达方式，能够传递很多文字和声音所不能表达的信息。人脸表情识别研究有非常重要的社会和经济意义，目前，表情识别在电子游戏、智能广告投放、智能人机交互、远程教育、安全驾驶、临床医学、幼儿教育与护理、心理学研究、智能监控、图像合成、动漫等方面都有广泛的应用，，且前景非常可观。近年来，随着表情识别的应用场景日渐推广和人们对表情识别领域的认识加深，表情识别技术已成为多个领域的热点研究课题。目前人脸表情识别的研究主要集中在表情特征提取和表情分类算法研究方面。本文针对表情图像序列，提出了基于AAM(ActiveAppearance Model)模型结合LK(Lucas-Kanade)光流跟踪算法的序列表情特征提取方法和基于随机森林的表情分类方法：首先，采用AAM模型对中性表情图像进行特征点定位，在后续图像帧间采用LK光流跟踪算法跟踪AAM人脸特征点；其次，将剧烈表情和中性表情图像中相应的AAM特征点的位移作为表情特征，采用SVM算法训练分类器，进行人脸表情运动单元分类；最后将人脸表情运动单元作为随机森林的输入进行训练得到表情识别分类器，进而对七种基本表情进行识别。本课题在Extended Cohn-Kanade人脸图像序列数据库进行了大量实验，实验结果表明：AAM模型结合LK跟踪算法的人脸序列特征提取算法较单纯的AAM人脸表特征提取更加精确高效；将图像序列中终止帧和起始帧之间AAM特征点的位移作为输入，采用SVM进行表情运动单元(Action Unit，简称AU)识别的识别率高达98%以上；采用人脸表情运动单元作为表情特征，随机森林算法的表情识别率是97.10%，而相同条件下贝叶斯网络的表情识别率为89.37%。另外，随机森林算法在训练和检测过程中都比贝叶斯网络快速高效。随机森林算法相比目前广为采用的贝叶斯表情分类算法，在表情识别率和算法效率方面都有很大提高。
[Abstract]:Facial expression is a very important way to express information in the process of interpersonal communication. It can convey many messages that can not be expressed by words and sounds.The research of facial expression recognition has very important social and economic significance. At present, facial expression recognition is used in video games, intelligent advertising, intelligent human-computer interaction, distance education, safe driving, clinical medicine, early childhood education and nursing.Psychological research, intelligent monitoring, image synthesis, animation and other aspects have a wide range of applications, and the prospects are very impressive.In recent years, expression recognition technology has become a hot research topic in many fields with the increasing application of expression recognition scene and the deepening of people's understanding of the expression recognition field.At present, the research of facial expression recognition mainly focuses on facial expression feature extraction and facial expression classification algorithm.In this paper, an expression feature extraction method based on AAM(ActiveAppearance Model model combined with LKG Lucas-Kanade-based optical flow tracking algorithm and an expression classification method based on random forest are proposed.AAM model is used to locate the neutral facial expression image, and LK optical flow tracking algorithm is used to track the AAM facial feature points between the subsequent image frames. Secondly, the LK optical flow tracking algorithm is used to track the AAM facial feature points.The displacement of the corresponding AAM feature points in the violent expression and neutral facial expression image is taken as the expression feature, and the classifier is trained by SVM algorithm to classify the facial expression motion unit.Finally, the facial expression motion unit is trained as the input of the random forest to obtain the facial expression recognition classifier, and then the seven basic expressions are recognized.A lot of experiments have been done in the Extended Cohn-Kanade face image sequence database. The experimental results show that the feature extraction algorithm based on the Extended Cohn-Kanade model combined with LK tracking algorithm is more accurate and efficient than the simple AAM face table feature extraction algorithm.The displacement of the AAM feature points between the termination frame and the start frame in the image sequence is taken as input, and the recognition rate of the facial expression motion unit (AAM) is up to 98% by using SVM, and the facial expression motion unit is used as the facial expression feature.The expression recognition rate of stochastic forest algorithm is 97.10, while that of Bayesian network is 89.37 under the same conditions.In addition, stochastic forest algorithm is faster and more efficient than Bayesian network in training and detection.Compared with Bayesian expression classification algorithm, stochastic forest algorithm has a great improvement in expression recognition rate and algorithm efficiency.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】