基于PCA和KHM聚类的唇特征提取算法的研究  

Lip Feature Extracting Research Based on PCA and KHM Clustering Algorithm

作　　者：刘颖[1] 王成儒[1] 

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004

出　　处：《微电子学与计算机》2008年第8期84-87,共4页Microelectronics & Computer

摘　　要：在基于隐马尔可夫模型(HMM)的语音同步唇动合成系统中,提出了基于主成分分析(PCA)法的口形全局纹理特征的表达方法,并在此基础上运用对初始值不敏感的K调和均值(KHM)聚类算法对实验中所用到的1650张24位真彩色唇图像样本进行PCA特征提取和聚类分析,以便于在语音唇动合成研究中建立语音特征与唇特征之间的映射关系.实验结果表明该方法保存了唇动过程中的嘴部二维形状信息,避免了由于特征点提取不准而对唇动合成产生的影响,同时运用KHM聚类算法使得聚类结果更加准确,有利于人脸语音动画的研究.This paper presents a global lip texture description based on principle component analysis for speech synchronous lip movement synthesize system based on Hidden Markov Models（HMM）. And with this basement, K harmonic means which isn＇t sensitive for the initialization is used for clustering 1650 pieces of 24 bit lip images in the research, in order to establish mapping relative between speech features and lip features during speech lip synthesizing research. The experiment indicates that the ideas presented in this paper not only keep lip texture during lip movement, but also avoid the bad influence in lip synthesizing because of the feature points unfaithfully extracting. KHM clustering algorithm makes cluster results steady with different initializations, clustering result becomes more accurate, so the ideas are proper for the research of facial speech animation.

关 键 词：特征提取 主成分分析 KHM 入脸语音动画 

分 类 号：TN911.73[电子电信—通信与信息系统]