权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

STUDY ON ABSTRACTION OF MOUTH SHAPE AND DEVELOPMENT OF ALGORITHM FOR QUICK RECOGNITION OF VOWELS TO REALIZE LIP READING IN ENGINEERING

工程中口型抽象及元音快速识别算法开发实现唇读

基本信息

批准号：
07650314
负责人：
KUROSU Kenji
金额：
$ 1.41万
依托单位：
KINKI UNIVERSITY
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
1995
资助国家：
日本
起止时间：
1995 至 1997
项目状态：
已结题

项目摘要

To realize lip reading by image processing, It is necessary to get the mouth shape correctly. As the means to do this, one approach is getting the front image of the mouth by using visual-servo which follows the movement of a face arid another approach is correcting the distortion caused by the angle of a camera to make pattern-matching with the standard pattern easy.In the first rind second year, visual servo which follows the movement of a face by using a robot with a camera in hand. Controller by neural network which follows the one or two dimensional marker are designed and tested experimentally. The main results are as follows :Neural network is composed of three layers which include 6 input neurons, 20 neurons in the second layer and 5 neurons in the output layer. The input signals are three feature points of an initial picture and an object picture, respectively. The manipulating variables for five joint angles are selected as the output signals and BP algorithm are used in network-learning to make the error of positioning zero . By this method, we can move the initial position of a camera to the target position by one step, if the Initial error is not so large.In the 3rd year, the correction of distortion caused by an angle of a camera is done using the 3rd order polynomials, whose parameters are determined by the least mean square error method. The matching with the standard pattern is performed using by the fuzzy degree of similarity and neural network and the results are successful in case of the same person. The correction method of the 3rd order polynomials is efficient for the correction of distortion in the inspection picture of Inside of pipes. So, application experiments for construction of panoramic pictures of Inside of pipes is reported as the paper on the journal.Moreover, survey of researches about lip reading is outlined.

要通过图像处理实现唇形识别，就必须准确地获取嘴型。一种方法是通过跟踪人脸运动的视觉伺服来获取口腔正面图像，另一种方法是校正摄像机角度造成的失真，以便与标准图案匹配。在第一年和第二年，视觉伺服通过手持摄像头的机器人跟踪人脸的运动。设计了跟踪一维或二维标记物的神经网络控制器，并进行了实验验证。主要研究结果如下：神经网络由三层组成，包括6个输入神经元、20个第二层神经元和5个输出层神经元。输入信号分别是初始图像和目标图像的三个特征点。选取五个关节角度的操作变量作为输出信号，采用BP算法进行网络学习，使定位误差为零。该方法在初始误差不太大的情况下，将摄像机的初始位置向目标位置移动一步；在第三年，利用最小均方误差方法确定参数的三次多项式，对摄像机角度引起的失真进行校正。利用模糊相似度和神经网络对标准模式进行匹配，在同一人的情况下取得了较好的匹配效果。三次多项式的校正方法对于管道内壁检测图像的失真校正是有效的。因此，本文报道了构建管道内部全景图的应用实验，并概述了唇读的研究概况。