基于特征层与分数层的多模态生物特征融合

发布时间：2018-01-02 04:28

本文关键词：基于特征层与分数层的多模态生物特征融合　出处：《中国科学技术大学》2017年博士论文　论文类型：学位论文

【摘要】：生物特征识别在人类的生产活动中,起到了越来越重要的作用。同时,随着生物特征识别越来越广泛的应用,单模态生物特征识别不足以满足实际使用中的诸多需求这一问题越发地突显出来。而多模态生物特征融合方法能够通过集成多个生物模态的信息,来缓解单模态生物特征识别系统在使用中所遇到的障碍。在最近的十几年里,多模态生物特征识别取得了迅猛的发展,然而仍有部分关键问题有待得到进一步解决,例如提取多模态互补信息、缓解噪声影响以及利用深层次信息等问题。本文基于多模态生物特征的特征层与分数层数据,从上述实际问题出发,探索能够有效缓解上述问题的融合方法和策略,同时建立方便、准确、高效的多模态生物特征识别系统。本文中,主要工作和贡献如下:·在高通量的场景下,以提取多模态生物特征之间的互补信息为目的,本文提出了一种基于稀疏线性规划的异质特征选择模型。基于该模型的融合方法能够提取多模态互补信息、去除冗余信息、降低所使用的异质特征维度并且减少所需的计算以及存储消耗,进而实现快速、准确地对用户身份进行识别。该模型具有所选特征稀疏、泛化能力强等诸多优良特性。由于该融合方法能够充分提取特征之间的互补性、去除冗余性,因此该方法能够在保证融合后识别性能提升的前提下,降低计算复杂度以及融合的时间消耗。·针对远距离生物特征识别系统中,由于环境噪声的干扰,致使融合后识别性能下降的问题,我们提出了一种对噪声鲁棒的融合模型。在生物特征识别系统中,噪声的干扰往往在获取数据环节最为强烈,并且会影响后续流程,如特征提取、特征比对等流程,进而影响最终融合的识别结果。本文针对噪声对分数层数据产生的影响,提出了一种基于对噪声充分估计的鲁棒线性规划模型,最大限度的在分数层面上消除了噪声的影响。该方法采用特征的线性组合模式,因此消除噪声影响的同时,还保证了算法的效率,保证了系统的实用性。该方法可以在系统的保守性与鲁棒性之间取得较好的平衡。本文提出了针对新型鲁棒线性规划模型的快速数学求解方法,将高度非线性模型转化为线性规划模型进行求解,使得求解过程更加简便高效。同时本文还从概率的角度对所提出的模型进行了阐述,计算出了该模型最优解情况下分类错误率的上确界,具有较为深刻的物理意义。·在特征数量有限的情况下,提出了一种能够从比对分数中抽取隐含信息的分段线性分类器融合方法。用户的生物特征信息在生物特征识别系统中,经由特征采集、特征提取、特征比对等模块时,虽然所操作的数据具有了规范化、可操作性强等特点,但是其所承载的用户生物特征信息随着数据处理而减少。而且生物特征比对分数与决策值之间的关系是高度非线性的,但是在通过比对分数计算得出决策值时,大部分系统都是采用简单的线性融合分类器来模拟该高度非线性的关系,这进一步导致了更多细节信息的丢失。为了恢复这些隐藏在分数层中的用户生物特征信息,本文提出了一种分段线性分类器融合框架。该方法受到"非线性是绝对的,线性是相对的"观点启发,通过对特征空间离散化的方法,自适应地学习得到比对分数与决策值之间的非线性函数关系。本方法可看作是将低维分数映射到高维空间的映射函数,因此也可以从核函数的角度,作为一种自适应的核函数映射方法。本文还对分段线性分类器融合框架下三种性能优异的融合分类器进行了分析与说明,并指出了其分别所擅长的实际应用中的问题。由于人类的脸部区域生物模态丰富,同时便于采集,而且在实际使用中能够与用户良好的交互,较多的多模态生物特征融合系统采用对人脸、虹膜、眼周等生物模态进行融合,不仅提升系统的准确性,还提高系统的易用性。因此,本文中也采用对人脸区域的生物模态进行融合,来验证本文中所提出的方法的有效性,从实际需求出发,建立便捷、准确的多模态生物特征识别融合系统。综上,本文对现有多模态生物特征融合工作中存在的问题进行了分析和研究,针对不同问题提出了具有创新性的解决方法,工作具有开创性的意义,推动了多模态融合识别的进一步发展。
[Abstract]:Biometrics in the production of human activities, plays a more and more important role. At the same time, with the application of biometrics is more and more widely, single modal biometric is not enough to meet the many needs of the actual use of this problem more and more prominent. The multi modal Biometric Fusion Method by integrating multiple a biological modal information, to ease the single modal biometric system encountered in the use of obstacles. In the recent years, multimodal biometrics has achieved rapid development, but there are still some key issues need to be further resolved, such as the extraction of multimodal complementary information, alleviate noise and use the deep information problem. This paper characteristic layer and fractional layer data in multi biometrics based on starting from the actual problem, can effectively alleviate the problem of exploration The fusion method and strategy, and establish a convenient, accurate, multi biometrics system with high efficiency. In this paper, the main work and contributions are as follows: in high-throughput scenarios, to extract the complementary information between multi biometrics for the purpose, this paper presents a selection model of heterogeneous feature sparse linear programming based on the fusion method of the model can extract multi modal based on complementary information, remove redundant information and reduce the dimension of the heterogeneity and reduce the computation required and memory consumption, so as to realize the rapid and accurate recognition of the identity of the user. This model has the characteristics of the selected sparse, strong generalization ability and many other excellent characteristics. Because the fusion method can fully extract complementary features, remove redundancy, so this method can guarantee the fusion recognition performance, reduce the calculation The complexity and integration time consuming. For remote biometric identification systems, because of the ambient noise, resulting in the problem of declining performance after fusion, we propose a fusion of noise robust model. In a biometric recognition system, the noise in the data link is the most intense, and will affect the follow-up process, such as feature extraction, feature matching process, thereby affecting the final fusion recognition results. The influence of noise on score level data, gives a method for noise estimation fully robust linear programming model based on the maximum score in the level to eliminate the effect of noise. This method by using the linear combination of pattern features, thus eliminating the noise at the same time, but also ensure the efficiency of the algorithm, to ensure the practicality of the system. In this way the conservatism of the system and To achieve a better balance between robustness. This paper proposes a fast method for solving the mathematical model of linear programming model is robust, the highly nonlinear model into a linear programming model is solved, which makes the solving process more convenient and efficient. Meanwhile, from the view of the probability of the proposed model are described, and calculated the optimal model the solution under the condition of the classification error rate of the supremum, physical significance is more profound. In a limited number of features, put forward to extract hidden information from the comparison of fractional piecewise linear classifier fusion method. The user's biometric information in a biometric recognition system, through the characteristics of acquisition. Feature extraction, feature matching module, while the operation data is standardized, strong operability, but the user biometric information it carries with The data processing is reduced. And the biological characteristics on the relationship between the score and the decision value is highly nonlinear, but by comparing scores calculated decision value, most of the system is the use of simple linear classifier fusion to simulate the relationship between the highly nonlinear, which further leads to the loss of more details. In order to restore these the hidden user biometric information in a fractional layer, this paper proposes a piecewise linear classifier fusion framework. This method is nonlinear is absolute, linear is relative "view of inspiration, through the method of feature space discretization, adaptive learning nonlinear function relationship between alignment scores and decision value. The method can be regarded as the low dimension is mapped to the mapping function in high dimensional space, so it can be from the nuclear function, as a kind of adaptive kernel The function mapping method. The properties of three kinds of piecewise linear classifier fusion framework excellent fusion classifier are analyzed and explained, and points out its practical application are good at problems in the face region. Because of rich human biological modal, and to facilitate the acquisition, but in actual use can interact with the user good, a multimodal biometric fusion system using more of the face, iris, fusion eye and other biological mode, not only to enhance the accuracy of the system, but also improve the usability of the system. Therefore, this paper also uses the biological modal of the face region of fusion, to verify the effectiveness of the method proposed in the paper. Starting from the actual demand, the establishment of convenient, accurate Multimodal Biometric Fusion System. In conclusion, the existing multi modal Biometric Fusion problems existing in the work of Analysis and research put forward innovative solutions for different problems. The work has pioneering significance and promoted the further development of multi-modal fusion recognition.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】