Adaptive multi-classifier systems for biometric recognition

对个人的生物特征识别为目前应用于许多安全和监视系统的传统认证方案提供了一种强大的替代方案。在实践中，生物识别系统的性能通常会下降，因为它们面临着在操作过程中发生变化的复杂环境，并且它们是使用有限的数据和底层数据分布的知识来先验设计的。生物特征模型往往不能很好地代表待识别的生物特征。为了准确识别，这些模型应该随着时间的推移而调整，以响应新的或变化的输入特征、数据样本、先验、类别和环境。该研究计划旨在研究自适应多分类器系统(AMCS)，该系统能够在真实世界的生物识别应用中实现高水平的性能，并有效地更新生物识别模型以响应来自操作环境的新兴信息。这些AMCS进化每个个体的二进制分类器(EoC)的集成，其中分类器使用基于种群的进化优化来联合训练。在个体加入系统的过程中，一种基于粒子群算法的动态多目标训练策略通过对数据样本的批量学习来生成多样化的基分类库。然后，作为对该个体的新数据的响应，该策略或者生成额外的池以与先前学习的分类器组合，或者通过增量学习来进化先前学习的分类器池。然后，根据准确度和多样性的专门衡量标准，从个人的池中选择分类器的子集。新的增量布尔组合技术被用来随时间调整决策级融合函数，以响应新的或变化的池。为了应对有限的数据和不对称的分布，在ROC或其他空间中应用了增量BC。虽然该方案中描述的稳健自适应技术可以应用于广泛的应用，但人脸识别、签名验证和生物特征融合是本研究的重点。为了加快这一计划的所有步骤，新的AMCS将在高速GPGPU平台上使用真实的生物识别数据进行验证。

Biometric recognition of individuals provides a powerful alternative to traditional authentication schemes that are presently applied in many security and surveillance systems. In practice, the performance of biometric systems typically declines because they face complex environments that change during operations, and they are designed a priori using limited data and knowledge of underlying data distributions. Biometric models are often poor representatives of the biometric trait to be recognized. For accurate recognition, these models should be adapted over time in response to new or changing input features, data samples, priors, classes and environments. This research program seeks to investigate adaptive multi-classifier systems (AMCSs) that can achieve a high level of performance in real-world biometric applications, and efficiently update biometric models in response to emerging information from the operational environment. These AMCSs evolve an ensemble of binary classifiers (EoCs) per individual, where classifiers are co-jointly trained using population-based evolutionary optimization. During the enrolment of an individual to system, a new dynamic multi-objective PSO-based training strategy generates a diversified pool of base classifiers through batch learning of data samples. Then, in response to new data for that individual, this strategy either generates an additional pool for combination with previously-learned classifiers, or evolves the pool of previously-learned classifiers through incremental learning. A subset of classifiers is then selected from an individual's pool according to specialized measures of accuracy and diversity. New incremental Boolean combination techniques are employed to adapt decision-level fusion functions over time, in response to new or changing pools. To account for limited data and skewed distributions, incremental BC is applied in ROC or other spaces. Although the robust adaptive techniques described in this proposal can be applied to a wide range of applications, face recognition, signature verification and biometric fusion are the focus of this research. To accelerate all steps of this program, new AMCSs will be validated with real biometric data on high-speed GPGPU platforms.