基于几何特征的人脸识别
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
基于几何特征的人脸识别
学号:姓名:
(上海大学机电工程与自动化学院,200072)
摘要:人脸识别技术作为多学科领域的、具有挑战性的课题,它覆盖了数字图像处理、模式识别、神经网络、数学等诸多学科的内容,同时也具有十分广泛的应用价值。在人脸识别领域,基于几何特征的人脸识别算法因其计算简单、使用有效等特性,引起了人们的广泛注意,并已成为人脸图像特征提取和识别的主流方法之一。本文定位人脸器官,通过人脸面部拓扑结构几何关系的先验知识,利用基于灰度投影的方法在知识的层次上提取人脸面部主要器官特征,将人脸特征用一组几何特征向量表示,识别归结为特征向量之间的匹配。本文工作包括:
(1)对灰度积分投影理论进行了详细的介绍和分析。这种方法是目前定位人脸轮廓的主要方法。在此基础上对一种新的在竖直方向上定位人脸左右轮廓的灰度差投影法进行了改进。投影法本质上是一种基于统计的方法,在具体应用时又结合了人脸特征分布的先验知识。这种方法不需要对积分投影图做任何平滑处理等操作,因而算法简单,准确率高,速度很快。
(2)提出了一种精确定位眼睛的方法,该算法将眼区灰度总体分布特点与眼部灰度变化特点相结合,将传统的积分投影法与灰度差累加值投影法相结合,通过大量试验选取合适的参数。试验结果表明,该算法对光照变化不敏感,定位准确率高。运用灰度积分投影结合人脸特征的先验知识定位鼻子,这种定位方法得到的准确率也是比较高的。嘴巴的定位则利用投影法求得。
(3)本文根据定位出来的人脸器官选出七个特征点,即四个眼角点、鼻尖点和两个嘴角点。利用它们构造了十个特征向量,并对其进行归一化运算。图像识别的最后一个过程就是分类,本文采取根据模式相似性的最近距离分类器进行分类。用加权比值函数来计算特征相似度,更适用于人脸图像的识别和计算。如何选择出合适的识别门限是个很复杂的问题,还有待于进一步研究。
关键词:人脸识别;灰度投影;几何特征;特征提取
The Research of Face Recognition Algorithm Based on
Geometric Features
Student number:15721637 Name:Zhao Pei-pei
(Institute of electrical and mechanical engineering and automation, Shanghai University, Shanghai 200072, China)
Abstract:the technology of human face recognition as a multi-disciplinary field and challenging psubject which contains digital image processing,pattern recognition,computer vision,neural network,psychology,physiology,mathematics and a good many subjects.In the meantime,it also has widely used.In the field of face recognition,the method of human face recognition based on geometric features has been paid great attention for its simple calculation and availability.At
present,it has become one of the dominant methods as the feature extraction and recognition.TMs article locates human face organs,through apriori knowledge of human face topological structure geometrical relationship,making use of method based on construct to extract the features of human face organs,expressing human face through a set of geometric feature vectors.The recognition putting in summary is matched with feature vector.This paper includes the following parts:
(1)Have a detailed introduction and analysis about the theory of greyscale integrated
projection.This method is now the main method of locating human face.We put forward a new method called greyscale differential projection which is based on the previous method and locating the contour of human face vertical directly.Projection method is essentially
based on statistics.It combines the apriori knowledge of human face feature distribution in the application.This method needn’t to do any pretreatment to the image and any smoothing treatment to the integrated projection image.So this algorithm is simple;the accuracy is high; the speed is quick.
(2)Give an introduction about the method of locating eyes precisely.This algorithm combines
the character of the eye area greyscale totally distribution and greyscale transformation;combines the methods of traditional integrated projection and differential projection.The experiment led to the fact that this algorithm is not sensitive to the illumination transformation and has a high accuracy.Using greyscale integrated projection combines the apriori knowledge of human face character to locate nose.This location method also has high accuracy.The location of mouth is abtained through projection method.
(3)The choice of characteristic points needs enough information and can’t go so far as to
increase calculation quantity.This article chooses seven characteristic points,namely,four canthus points,tip of nose and two corners of mouth points.Construct ten eigenvectors using them and carries on the normalization calculation to them.The last process of image recognition is classification.After adopting some standards to extract feature of human images,we construct category separability decision rule according to these characters and design classifier.This article takes use of minimum distance classification to classify.It iS more suitable for human face recognition and calculation using weighing ratio to calculate similarity.How to choose a suitable recognition threshold is a difficult problem and need further research.This article ascertains it through a good many experiments.
Key Words:Face Recognition;Greyscale Projection;Geometric Characters;Feature Extraction
1. 引言
1.1人脸识别技术: