人机交互中视觉语言的灰度轮廓权向量差分唇形特征模型  

A Model of Intensity Profile Weight Vector Differential Shape Feature in Visual Languages for Human-Machine Interaction

作　　者：甘俊英[1] 赵向阳[2] 张有为[3] 

机构地区：[1]北京航空航天大学电子工程系,北京100083 [2]中国科学院声学研究所,北京100080 [3]五邑大学信息科学研究所,广东江门529020

出　　处：《计算机工程与应用》2002年第13期17-19,52,共4页Computer Engineering and Applications

基　　金：广东省自然科学基金(编号:000872)

摘　　要：该文结合函数变形模型和灰度轮廓向量模型的特点,给出了一种维数少、有效性高的视觉语言特征—灰度轮廓权向量差分唇形特征模型。该特征融合了嘴唇图像的形状变化信息和灰度信息,能够较完善地描述嘴唇的变化。同时,得出了一种新的视觉特征提取算法。仿真结果表明,该算法与传统的函数变形模型相比,总的特征提取准确率提高了5个百分点,每个发音图像序列特征提取的准确率提高了1.6至9个百分点,每帧图像的特征提取时间由4.6495秒下降到0.4455秒。对“1”至“10”数字发音的嘴唇图像序列进行识别,获得了较高的识别率。因此,灰度轮廓权向量差分唇形特征是一种精炼、描述性强、适合于唇读识别的视觉语言特征,该算法能自动地完成模型的训练和视觉特征的提取,是一种有效的特征提取算法。Combined with the characteristics of function deforma bl e model and intensity profile vector model,a sort of visual language,named int ensity profile weight vector differential shape feature,is presented in this p aper,whose dimension is small and validity is high.It can perfectly describe l ip change because of the fusion information of lip change and intensity change. In the meantime ,a new visual feature extraction algorithm is derived.Simulati on results demonstrate that compared with traditional function deformable model ,the total feature extraction accuracy rate is improved5%,the feature extrac tion accuracy rate of each pronunciation image series is improved1.6%to9% ,and the feature extraction time of each frame image is reduced from4.64 95s to0.4455s.A higher recognition rate is obtained when used in lip im age series of digit pronunciation from1to10.Therefore,this feature is a vis ual language feature proper for lipreading,while the algorithm is valid in mode l training and feature extraction.

关 键 词：人机交互 视觉语言 灰度轮廓 权向量差 分唇形特征模型 语音识别 图像识别 

分 类 号：TN912.34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]