【作者】 李永；

【导师】 殷建平；

【作者基本信息】 国防科学技术大学 ， 计算机科学与技术， 2011， 博士

【摘要】 受数据噪音和识别系统本身的限制,基于单一生物特征的身份认证系统所能达到的准确率是有限的,通过多生物特征识别来提高识别的准确率成为当前生物特征识别领域的研究热点之一。本文结合具体生物特征识别,尤其是指纹识别,从多样本特征模板选择、基于模拟数据和真实数据的多单元识别、多角度非接触指纹识别、多模态识别等研究了基于匹配分数的多生物特征识别,试图为多生物特征识别提供可行的解决方案。本文的主要工作和创新点如下：1.针对多样本模板选择问题,提出了最大化匹配分数算法MMS(Maximum Match Scores)和贪婪最大化匹配分数算法GMMS(Greedy Maximum Match Scores)。MMS和GMMS算法不需要了解原始生物特征数据的细节,避免了深入特征提取过程,因此该方法更加灵活,能够应用于各种生物特征识别的模板选择过程。将本文的算法与已有的算法进行比较可知,通过MMS和GMMS算法进行特征模板选择可以有效提高生物特征识别系统的识别性能。针对模板更新问题,提出了两种策略,在线策略和离线策略,这两种策略各有优缺点,实验表明,离线策略能够取得更好的识别性能。2.针对多单元识别系统,提出了一种基于均方根的多单元融合规则Square。本文首先针对多单元指纹识别,基于模拟多单元数据比较了不同融合规则的识别准确率,然后提出了一个新的融合规则Square。基于模拟数据的实验表明,在FAR较小和较大时,Square和Sum规则能够分别获得最优识别性能。结合Square和Sum规则的优点,设计了一种新的规则Square-Sum,通过理论分析和实验证明,Square-Sum算法可以取得更优的识别性能。3.研究了真实十指指纹的单指识别性能、多指融合性能和指间相关性,提出了一种改进的Sum规则。本文以MCYT-Fingerprint真实十指指纹数据为基础,进行了大量实验,并结合前人的研究,对单个指纹的识别性能进行了比较,分析了多个手指融合的识别性能和不同手指之间相关性。同时,针对真实十指指纹数据,提出了一种改进的Sum规则,实验表明该规则在真实数据上有更好的识别性能。4.针对多角度非接触指纹识别,提出了一种基于聚类的动态分数选择算法CDSS。采集了一个多角度非接触指纹数据库,研究了非接触指纹的预处理、特征提取和特征匹配过程,初步实现了非接触指纹识别系统。针对多角度非接触指纹识别,提出了基于聚类的动态分数选择算法,通过聚类获得了匹配分数之间一些新的信息,根据这些新的特征以及匹配分数一些统计量的数值大小关系,灵活地选择不同融合规则作为最终的融合策略。相对于单个角度指纹识别,多角度非接触指纹识别可以显著提高系统的识别性能。同时,与Sum、Max、SVM和Fisher线性判别算法的实验比较表明,CDSS算法可以获得更好的识别性能。5.针对多模态识别,提出了基于FRR和FAR的融合算法。首先提出了一种基于FRR和FAR信任度融合算法。该融合算法以FRR和FAR为基础,既避免了直接求取某个匹配分数的后验概率,同时又能够刻画匹配分数的分布。通过改进信任度算法,提出了一种基于SVM的FRR和FAR融合算法。算法实现过程中,首先计算训练集中出现的每个分数的FRR和FAR值,然后通过插值的方法来计算测试集上出现的匹配分数的FRR和FAR值。在寻找插值的左右节点时,采用了计算量较小的二分查找法。最后通过实验验证了本文的算法具有良好的识别性能。更多还原

【Abstract】 Influenced by data noise and limitation of recognition system itself, the accuracy of identification system based on single biometric trait proves to be quite limited. Therefore, the reaserch of using multi-biometric recognition for improving recognition accuracy has become one of the hotspots in biometric recognition field. Combining with specific biometric recognition, especially fingerprint recognition, this paper studies multi-biometric recognition based on matching score from the aspects of multi-sample feature template selection, simulated data and real data based multi-unit recognition, multi-view touchless fingerprint recognition, multi-modal recognition and etc., aiming at providing possible solution to problems in multi-biometric recognition process. The main work and contributions are as follows:1. For multi-sample template selection, this paper proposes two algorithms, MMS (Maximum Match Scores) and GMMS (Greedy Maximum Match Scores). MMS and GMMS algorithms do not involve details of primitive biometric feature data and eliminate the necessity of thorough feature extrcation process. Therefore, the two algorithms are more flexible and can be used in various biometric systems. They prove to be effective for template selection compared with the existing algorithms. For template updating, we put forward two methods, online and offline methods. Each of the two strategies has its own advantages and disadvantages. The experimental results show that offline strategy can get better performance.2. For multi-unit recognition system, this paper proposes a fusion algorithm Square. For multi-unit recognition system, this paper first compares identification accuracy of different fusion rules based on pseudo multi-unit data, and then put forward a new fusion rule:Square. According to our simulated data experimental results, Square and Sum rule could respectively achieve the best recognition performance when FAR is smaller and bigger. We then design, based on the advantages of Square and Sum rule, a new Square-Sum rule, which could always get better recognition performance.3. Based on real multi-unit fingerprint data, this paper studies the recognition performance of single finger, performance of multi-finger fusion, correlation between different fingers, and proposes an improved Sum rule for multi-unit biometric systems. Through a number of experiments based on MCYT-Fingerprint dataset and with the combination of previous studies, this paper compares the recognition performance of single fingerprint, analyzes recognition performance of multi-finger fusion, studies the relativity of different fingers, and proposes an improved Sum rule for multi-unit biometric systems as well as verifying that the improved Sum can gain better performance.4. For multi-view touchless fingerprint recognition, this paper proposes a Cluster-based Dynamic Score Selection algorithm CDSS. We collect a small-scale multi-view touchless fingerprint database, study the image preprocess, feature extraction and feature matching of touchless fingerprint, and basically implement touchless fingerprint recognition. This thesis studies multi-view touchless fingerprint recognition based on matching score fusion, then puts forward CDSS multi-view touchless fingerprint recognition fusion method. After obtaining the new features of different matching score through clustering, we flexibly select different fusion rules as the final fusion strategy on the basis of these new features and statistics value size comparison of matching score. Compared with single-view fingerprint recognition, multi-view touchless recognition could greatly improve the recognition performance of the system. At the same time, CDSS is compared with Sum, Max, SVM and Fisher linear discriminate, which proves the method of this paper gets better recognition performance.5. For multi-model recognition, this paper proposes fusion algorithms based on FAR and FRR. This paper firstly proposes a new confidence-based fusion strategy based on FAR and FRR. As confidence-based strategy is established on the basis of FRR and FAR, it could both avoid directly accessing the posterior probability of certain scores, and dipict the distribution of scores. Then we improve the confidence-based algorithm and propose a fusion algorithm using SVM based on FRR and FAR. This algorithm could utilize both some overall situation information, that is the corresponding FAR and FRR value of matching score, and the good classification capacity of SVM. During implementation, this paper first calculates the transformation value of score having appeared in training set, and taking these values as fixed nodes, calculates the transformation value of matching scores having appeared in testing set by method of interpolation. When looking for the left and right node for interpolating value, this paper adopts binary searching strategy which greatly reduces computation. The experimental results show that the algorithms could achieve better recognition performance.更多还原