【作者】 周斌；

【导师】 贾惠波；

【作者基本信息】 清华大学 ， 仪器科学与技术， 2007， 博士

【摘要】 多生物特征融合是生物特征识别领域里的一个重要课题,它分为采集层、特征层、量化层和决策层四个融合层次,本论文研究的重点是量化层融合。本论文从统计分布的角度出发,对量化层融合进行研究。通过理论研究和实验证明了量化层融合可以有效的提高系统识别性能,并得到了最佳融合策略选择的方法。本论文首先在已有量化值归一化模型的基础上,提出了一种新的量化值归一化模型——双期望统一模型(DUE)。相比其他量化值归一化模型,这一模型能够进一步简化不同特征之间量化值集合的分布关系。然后在这一模型的基础上,对平均加和法(SA)、加权平均法(WS)、最小值法(MIS)和最大值法(MAS)四种量化层的融合策略进行了理论研究。研究的内容包括,量化层融合后的分布情况,融合前的量化值分布对系统识别性能的影响,不同条件下最佳融合策略的选择等问题。在研究的过程中得到了一系列结论,这些结论将在后面多生物特征融合实验中得到验证。为了提供多生物特征融合的实验样本,本论文研究了两种新兴的具有广阔应用前景的生物特征——掌纹和手背血管。针对掌纹特征,提出了编码特征和纹理分布特征的提取算法;针对手背血管特征,提出了细节结构特征和纹理信息编码特征的提取算法。以此为基础,进行了三组多生物特征融合实验,掌纹与手背血管融合实验,掌纹编码特征与纹理分布特征融合实验和左右手掌纹融合实验。实验结果充分证明了量化层融合可以有效的提高系统识别性能。描述整体系统识别性能的参数,等错率EER在三个实验中融合后都显著下降。量化层融合策略的理论分析中得到的一系列结论在实验中也得到了验证,实验结果基本符合。通过对可能产生误差的因素进行的分析,本论文的模型基础和理论推导的结论在实际工程应用中具有较强的可行性和鲁棒性。利用本论文对量化层融合理论研究的成果,可以根据融合前单一生物特征的分布,预估多生物特征融合系统的识别性能,并可以根据实际需求确定最终的融合策略。因此本论文的研究对多生物特征融合系统的构建和应用具有实际意义。更多还原

【Abstract】 Multimodal biometrics fusion can effectively improve the performance of a unimodal biometric identification system. Fusion in multimodal biometric systems can take place at four major levels: sensor level, feature level, score level and decision level. From a statistical point of view, this thesis focuses on score level and its effect on system recognition.Herein a new score normalized modal, double uniform expectation modal (DUE), is proposed. Compared to other normalization modals, this modal can simplify the relation among score distributions from different features. Furthermore, based on this modal, four score level fusion strategies (Sum-Average, Weight-Sum, Max-Score and Min-Score) are studied. The study includes distribution after fusion, effect to recognition performance by score distribution before fusion, and selection of optimal fusion method under different conditions. During research, several conclusions are gained and are testified in experiments of multimodal fusion.To provide multimodal biometrics fusion samples, this thesis refers to palmprint and hand vein pattern. Regarding palmprint, extraction algorithm of palmprint-code feature and texture distribution feature are proposed; With regard to hand vein pattern, extraction algorithm of minutia structure feature and information code feature are proposed. Using above features, three experiments are performed. These experiments sufficiently prove that score fusion can improve system recognition effectively, and equal error rate (EER), an important parameter to describe system recognition performance, decreased distinctively. The conclusions from theoretical analyses are proved by these experiments, and accord with experiments results. By analyzing factors which induce errors, the modal and the theoretical deduction are proved to have strong feasibility and robusticity.Using the results of theory research on score level fusion, we can estimate the performance of multimodal biometrics fusion system by analyzing the distribution of unimodal biometric, and establish the optimal fusion strategy according to certain demand, therefore this thesis means much to the construction and application of multimodal biometrics system.更多还原