第6卷第3期 智能系统学报 Vol.6 No.3 2011年6月 CAAI Transactions on Intelligent Systems Jun.2011 doi:10.3969/j.issn.16734785.2011.03.007 基于局部纹理ASM模型的人脸表情识别 彭程,刘帅师1,万川,田彦涛2 (1.吉林大学通信工程学院,吉林长春130025:2.吉林大学工程仿生教育部重点实验室,吉林长春130025) 摘要:针对主动形状模型(ASM)迭代过程容易陷入局部最优解的不足,提出了一种基于局部纹理模型的改进ASM 算法,即EWASM.在局部纹理模型构建中,以每个特征点的中垂线方向搜索其邻域信息以确定最佳匹配位置,对衡 量匹配程度的马氏距离加以推广,进而得到改进的扩展加权局部纹理模型,它由中心局部纹理模型、前局部纹理模 型和后局部纹理模型共3个子模型加权组成,并对加权参数进行实验优化,使各个特征点之间的联系更加紧密,模 型的鲁棒性更好.通过表情识别实验对提出的EWASM算法和传统ASM算法进行对比,选用RBF神经网络分类器进 行表情分类,实验结果表明EWASM算法收敛速度更快,识别率也得以提高,并解决了局部最小问题,能更有效地表 征表情. 关键词:人脸表情识别:主动形状模型:局部纹理模型;RBF神经网络分类器 中图分类号:TP391文献标识码:A文章编号:16734785(2011)03023108 An active shape model for facial expression recognition based on a local texture model PENG Cheng',LIU Shuaishi',WAN Chuan',TIAN Yantao1.2 (1.School of Communication Engineering,Jilin University,Changchun 130025,China;2.Key Laboratory of Bionic Engineering (Ji- lin University),Ministry of Education,Changchun 130025,China) Abstract:An improved active shape model(ASM)called EWASM (expanded weighted ASM)based on a local tex- ture model was proposed because EWASM overcomes the disadvantage that the active shape model is easy to involve in local optimal solution in the iterative process.In the local texture model,searching adjacent information of each landmark along its perpendicular bisector made the match position best.It improved and promoted Mahalanobis dis- tance which measured the matching degree.Then the local texture model was extended to include the center local texture model,forward local texture model,and backward local texture model.After that,the weighted parameters were optimized experimentally.Thus each landmark is more closely related and the local texture model is more ro- bust.Finally facial expression recognition experiments were conducted comparing EWASM with classical ASM,and a RBF neural network was used as a classification in the expression recognition.Experiments show that the EWASM algorithm solved the local minimum problem and achieved a better convergence rate and recognition effect. Keywords:facial expression recognition;active shape model;local texture model;RBF neural network classifier 人脸表情不仅在人与人的交流中发挥着重要的中是一个关键环节.目前最受关注的特征提取方法是 作用,而且是实现人机交互,使计算机能够更准确地 Cootes等人于1995年提出的主动形状模型(active 理解人的表情和意图的一个重要研究内容.一个完整 shape model,ASM)方法12i,其模型允许一定程度上 的人脸表情识别系统由人脸检测、特征提取和表情分 形状的变化,可以更好地定位物体的内外轮廓,又不 类3个部分组成.其中特征提取在整个表情识别过程 会脱离目标对象的本质特征;但该模型对初始形状的 定位非常敏感,本质上是一个求局部最优的过程,有 收稿日期:2010-1126. 基金项目:吉林省科技发展计划重点资助项目(20071152):吉林大学 时不能发现全局最优解,存在着一定的局限性,因此 “985工程”工程仿生科技创新平台项目资助;吉林大学研 究生创新基金资助项目(20101027). 吸引了国内外广大学者对其进行研究, 通信作者:田彦涛.Email:tianyt@jlu.edu.cm. 1998年,Cootes等人进一步提出了主动表观模