【作者】 胡颖杰；

【导师】 蒙应杰；

【作者基本信息】 兰州大学 ， 计算机软件与理论， 2011， 硕士

【摘要】 随着传统安全认证以及网络社会的发展,安全有效的实时性身份认证越来越受到人们的重视,在基于生物特征的实时身份识别技术中,生物标识性动态特征的提取是决定算法识别效果、可行性的关键。本文针对现有研究大多基于单个静态图像,未能充分利用唇部变化特征的局限性,以及特征提取中需处理的信息量较大、识别率不高等缺陷,提出基于时间序列的动态唇形身份识别思想,在提取讲话人的特征向量时间序列的基础上,建立反映讲话人特征的隐马尔科夫模型,进而完成身份识别的判定。论文的主要研究工作和成果体现在以下几个方面。(1)在研究已有生物特征识别及唇部动态特征提取技术的基础上,构造了一种基于时间序列的动态唇形识别系统,设计了系统的体系结构,系统包括预处理、特征提取以及基于时间序列的身份识别三部分。(2)在特征提取部分,设计了基于点模型的内唇特征信息混合提取模型；给出了模型各组成部分的功能关系,设计了模型中基本唇印序列构造、平均模型计算、特征向量序列构造等过程的具体处理算法；通过仿真实验对基于内唇的混合特征提取算法的可行性进行了验证。(3)在识别部分,设计了基于时间序列的身份识别模型,利用隐马尔科夫模型来处理时间序列；设计并给出了隐马尔科夫模型状态确定、参数初始化、训练优化以及利用出现概率进行识别等几个关键环节的具体处理算法；通过仿真实验对基于时间序列的身份识别算法的性能进行了验证和对比分析。更多还原

【Abstract】 With the developments of traditional security authentication and network society, identification, which was effective, safe and real-time, got more and more attention. In the identification based on biological feature, the extraction of biological symbol feature decided the recognition effect..Because the traditional feature extraction of lips based on single static image neglected the movement, and the problems of huge information and low recognition rate, we proposed a method of speaker identification in time-sequential images based on movements of lips. This method extracted feature information of each frame in time-sequential of videos, and HMM models of movements of lips of speakers were constructed and trained to identify. We designed the whole system and the main researches were given below:(1) Based on the researches of some feature extraction based on viewable biologic feature, we constructed a identification system which used time-sequential images to capture the movement of lips to identify. This system comprised three parts:pretreatment, feature extraction and identification.(2) We designed a mixed extraction model with the feature information of internal lips based on feature points, and decreased dimensions of characteristic vector by calculating the similarities. The functional relationships of each part of model and the specific processing methods were given. At last, we verified the feasibility by simulation experiments.(3) We designed the identification model based on time-sequence, which built HMM models to generate observed symbol sequence, and did match with the models in database. The details of several key algorithms were given too, and the performances of our algorithms were analyzed by experiments.更多还原