【作者】 金怀国；

【导师】 孙冬梅；

【作者基本信息】 北京交通大学 ， 模式识别与智能系统， 2012， 硕士

【摘要】 随着信息技术的发展和人类社会的进步,信息安全成为一个越来越棘手的问题。传统的身份识别基本是基于身份标识的,如证件、密码、卡号等,但是这些却不能够真正的确认使用者是不是用户本人,一旦他人冒用,将会给持卡人带来不必要的损失。为了克服缺陷,基于人体生物特征的身份识别技术逐渐发展起来,比如指纹、掌纹、人脸、虹膜。特别是指纹识别,发展相对成熟,常常被各公司用于员工考勤。但是,由于指纹、掌纹都处在人体外部,很容易受到损伤,例如一些用手劳动的人,可能会磨损指纹、掌纹,久而久之,可能完全破坏指纹和掌纹的结构,导致注册不成功。因此,作为体内生物特征之一的静脉特征便成为首选。特别是,手指静脉识别以其采集器小,处理灵活而著称。本文以手指静脉识别技术为研究背景,重点研究了基于手指静脉特征点和子空间的特征匹配方法。主要工作有：1、制作了一个基于手指谷点定位的手指静脉图像采集装置,并建立了一个小型的手指静脉图像数据库,共计630幅图像,每个人30幅(食指、中指及无名指各10幅图像),共计21人。2、研究了手指静脉图像的预处理、静脉特征提取、后处理等工作。手指静脉图像的预处理主要包括感兴趣区域(ROI)定位、尺寸以及灰度归一化；在手指静脉特征提取这一部分,主要介绍了如何利用基于静脉纹路走向的谷形检测算法分割出手指静脉纹路；后处理的内容是对分割后的手指静脉纹路中离散点的去除以及孔洞的填充等工作。最后,对处理后的手指静脉纹路图像进行骨骼化,得到手指静脉纹路的骨架,并提取手指静脉特征点。3、在基于手指静脉特征点的特征匹配方法中,针对传统的仅仅利用正向平均豪斯道夫距离(FMHD)来刻画待匹配手指静脉图像到注册手指静脉图像静脉特征点间的距离,提出一种基于Fisher准则的融合了正向和反向平均豪斯道夫距离的新的特征点集间相似性的度量方式。实验结果表明,采用融合后的匹配分数进行特征匹配,识别率从54.81%提高到了87.03%。4、在基于子空间的特征匹配方法中,在详细介绍了主成分分析(PCA)、最大边界准则(MMC)、线性判别分析(LDA)、基本原理的基础上,实现了一维主成分分析(1DPCA)、一维最大边界准则(1DMMC)、一维线性判别分析(1DLDA)、二维主成分分析(2DPCA)、二维最大边界准则(2DMMC)、二维线性判别分析(2DLDA)以及2DPCA+2DPCA、2DMMC+2DMMC、2DLDA+2DLDA、2DPCA+2DMMC、2DPCA+2DLDA、2DMMC+2DLDA等方法提取手指静脉特征并进行手指静脉特征匹配。实验结果表明,2DPCA+2DLDA、2DMMC+2DLDA和2DLDA+2DLDA三种方法的识别率最高,而且还能够保证认证时间和压缩性能分别保持在1s和100维左右。更多还原

【Abstract】 With the development of information technology and the progress of human society, information security has become an increasingly thorny issue. Traditional identification is basically based on identity, such as documents, password, card number, but it can’t really verify that the user is not the user himself, once others falsely, will give the cardholder unnecessary losses. In order to overcome the defects, biometric identification technology gradually developed, such as fingerprints, palm prints, face, iris. Fingerprint identification, development is relatively mature, often used for employee attendance. However, because of the fingerprints, palm prints are in the human external, are susceptible to damage, such as hand labor, may be worn fingerprints, palm prints, with the passage of time may completely destroy the structure of fingerprints and palm prints, leading to registration is unsuccessful. Therefore, vein feature, as one of the in vivo biological features, become the first choice. Especially, finger vein recognition for its acquisition, flexible processing is known.In this dissertation, we are under the background of finger vein recognition technology, focusing on the feature matching methods based on finger vein feature points and subspace. The main tasks as follows:1、A new finger vein capturing device is designed, which has the valley points of the fingers fixed, and three finger vein (index finger、middle finger、ring finger) images database is built. There are a total of630finger vein images and21persons for each person30images (each finger10images).2、The work of finger vein image preprocessing, finger vein feature extraction and aft-processing was introduced. Finger vein image preprocessing contains region of interest (ROI) location, size and gray-scale normalized. In the part of the finger vein feature extraction, mainly introduces how to use finger vein pattern to the valley shape detection algorithm to split a finger vein patterns. Aft-processing is the work of wiping out the discrete points after the finger vein patterns divided and filling the void. Finally, skeletonize finger vein pattern image, get the skeleton of the finger vein patterns, and extract finger vein feature points.3、Due to feature matching method based on finger vein feature points, which only use Forward Mean Hausdorff Distance (FMHD) to characterize the distance from the matched finger vein feature points to the enrolled finger vein feature points. We propose a similarity measure between the new feature point sets based on Fisher Criterion, fusing the Forward and Reverse Mean Hausdorff Distance. The experiment results show that Recognition Rate (RR) respectively increases from54.81%to87.03%.4、In the part of feature matching method based on subspace, we introduces the basic theoretics of Principal Component Analysis (PCA), Maximum Margin Criterion (MMC), Linear Discriminat Analysis (LDA), then, use the method of One Dimensional Principal Component Analysis (1DPCA), Two Dimensional Principal Component Analysis (2DPCA), One Dimensional Maximum Margin Criterion (1DMMC), Two Dimensional Maximum Margin Criterion (2DMMC), One Dimensional Linear Discriminat Analysis (1DLDA), Two Dimensional Linear Discriminant Analysis (2DLDA) and2DPCA+2DPCA、2DMMC+2DMMC、2DLDA+2DLDA2DPCA+2DMMC、2DPCA+2DLDA、2DMMC+2DLDA to extract finger vein feature and match them. The experiment results show that Recognition Rate (RR) of three feature matching methods,2DPCA+2DLDA、2DMMC+2DLDA and2DLDA+2DLDA is the highest, but also they ensure that recognition time and compressive performance are maintained about1s and100dimensional space respectively.更多还原