基于鲁棒组件表达及组件关系的异构多生物特征融合研究

Multibiometrics has drawn much attention for its high accuracy, robustness, flexibility and security, which is reported as the inevitable trend of automatic personal identification technology. However, the diversity of biometric traits leads to the heterogeneous biometric representations in terms of data structure, data pattern, data scale, data dimension and the corresponding matching function. Because of this, most current fusion methods cannot effectively utilize the intrinsic information of each trait and the relationship among multiple traits, especially for the incompatible data structures and data patterns, which not only degrades the system performance but also increases both the computational cost and storage burden. Our previous work shows that there exists plenty of redundancy as well as complementarity among the local components of multiple biometric traits. Based on this, our project will start with the biometric representation based on pictorial model which divides the original biometric data into several components. A deformation factor is introduced to characterize rigid displacement of component to construct deformable pictorial model for biometric representation. Then, the large margin learning with hybrid norm regularization term is adopted to select the most distinct components. After that, a (dis)similarity measure based space transformation is utilized to transfer the heterogeneous representations of different biometric components into the same metric space, and the pictorial relationship is modeled in the metric space to construct the graph model for the components. At last, the graph embedding based manifold learning is adopted to discover the latent complementarity among the components, and reduce the computational cost and storage burden. To verify the effectiveness and efficiency of the proposed methods, the fusion of classical representations of several popular biometric traits is tested on the public databases and real applications. Our research provides a new solution and implementation of heterogeneous feature fusion for multibiometrics.

多生物特征识别由于精确度高、稳定性好、适应性强、安全性高被认为是自动身份识别技术的必然发展趋势。然而，生物特征物理模态的多样性造成了表达方式的异构性，导致现有特征融合方法缺乏对生物特征内部有效信息和多生物特征之间关系的挖掘，既制约系统识别性能的提高，又加重系统的计算和存储负担。现有工作和我们的前期工作表明，多生物特征局部模块表达鲁棒性强，且模块之间具有较强的冗余性和互补性。为此，本项目拟采用形变组件模型和基于稀疏表达的大边界学习方法构建鲁棒高效的生物特征表达；通过度量空间转换克服组件之间的异构性并构建多生物特征组件图模型；利用图嵌入流形学习挖掘组件之间的潜在互补性，实现异构多生物特征融合。在理论分析的基础上，整合现有数据资源建立验证平台，以常用的生物特征为研究对象，对本项目的方法进行测评。本课题的研究深入挖掘了多生物特征之间的互补关系，为异构特征融合提供了全新的解决思路和高效的实现方法。

多生物特征识别由于精确度高、稳定性好、适应性强、安全性高被认为是自动身份识别技术的必然发展趋势。然而，生物特征物理模态的多样性造成了表达方式的异构性，导致现有特征融合方法缺乏对生物特征内部有效信息和多生物特征之间关系的挖掘，既制约系统识别性能的提高，又加重系统的计算和存储负担。 现有工作和我们的前期工作表明，多生物特征局部模块表达鲁棒性强，且模块之间具有较强的冗余性和互补性。本项目提出了采用形变组件模型和基于稀疏表达的大边界学习方法构建鲁棒高效的生物特征表达；通过度量空间转换克服组件之间的异构性并构建多生物特征组件图模型；利用图嵌入流形学习挖掘组件之间的潜在互补性，实现异构多生物特征融合。本项目研究的基于光场成像的多生物特征图像采集技术，实现了在远距离实现对虹膜、人脸、掌纹数据的获取，为多生物特征融合提供了数据采集基础；本项目研究的基于词典学习的人脸图像组件模型特征表达方法取得了同类算法中的最优结果；本项目研究的基于相关熵的旋转不变最小二乘回归方法，为多生物特征的多维度特征选择提供了有效的特征选择方法；同时，本项目还研究了基于组件模型的人脸图像视频检索方法，开发了新闻视频的媒资检索系统软件，相关产品应用于某部门的综合声像信息系统。由于本项目的研究成果，项目负责人受邀参加了第四届中美工程前沿青年研讨会和第一届中印工程前沿青年领袖研讨会，并大会上进行了主题报告。本项目的研究为今后的多生物特征识别提供了新的思路，为拓宽多生物特征识别技术的工程应用提供了可行的方法。