基于不确定性信息距离的多分类器差异性度量构造与应用研究

As a decision-level information fusion theory and method, multiple classifier systems have been widely used in complicated pattern classification problem. However, using multiple classifiers doesn't necessarily improve the classification performance. Diversity among member classifiers is a prerequisite for effective fusion. Because traditional diversity measures in multiple classifier systems depict member classifiers with relatively coarse and ineffective feature information, there doesn't exist the obviously positive correlation between the traditional diversity measure and the fusion classification performance. Moreover, the diversity can even be drowned. The lack of effective diversity measure has restricted the development of multiple classifier systems. This project uses more abundant and delicate way for member classifier information modeling including output classes ranking, belief function, membership function, probability, and statistical histogram etc. Then information distances are designed based on the aforementioned uncertainty information for describing the diversity. In this project, the remote sensing terrain classification applications are researched to support the above theoretical researches. The innovative research achievement will enrich and perfect the multiple classifier systems, and can provide a solid theoretical foundation and technological reserve for the successful applications of the new achievements in the complicated pattern classification problems.

多分类器系统作为一种决策级信息融合理论与方法，在复杂模式分类问题中取得了重要应用。然而采用多个分类器并非总能带来性能提升，多分类器间的差异性是有效融合的前提和基础。传统多分类器系统差异性度量中用于描述个体分类器的特征信息较为粗糙、流于表象，因而传统度量与融合分类性能缺少明显的正相关关系，且存在诸如“差异性淹没”等缺陷。缺少合理有效的差异性度量已成为制约多分类器系统理论与方法发展的瓶颈。因此本项目研究采用更丰富、细致的方式对个体分类器信息进行建模，包括个体分类器输出类别排序、信度、隶属度、概率、统计直方图等信息，并基于上述不确定性信息设计相应的距离，进而刻画差异性。本项目将以遥感地物分类问题作为应用研究对象，以支撑上述理论层面的研究。力争取得一批原创研究成果，以期在丰富和完善多分类器系统理论与方法的工作中有所突破，并为新成果在复杂模式分类问题中得以成功应用，提供坚实的理论基础与关键技术储备。

多分类器系统作为一种决策级信息融合理论与方法，在复杂模式分类问题中取得了重要应用。然而采用多个分类器并非总能带来性能提升，多分类器间的差异性是有效融合的前提和基础。传统多分类器系统差异性度量中用于描述个体分类器的特征信息较为粗糙、流于表象，因而传统度量与融合分类性能缺少明显的正相关关系，且存在诸如“差异性淹没”等缺陷。缺少合理有效的差异性度量已成为制约多分类器系统理论与方法发展的瓶颈。因此本项目研究采用更丰富、细致的方式对个体分类器信息进行建模，包括个体分类器输出类别排序、信度、隶属度、概率、统计直方图等信息，并基于上述不确定性信息设计相应的距离，进而刻画差异性。本项目将以遥感地物分类问题作为应用研究对象，以支撑上述理论层面的研究。力争取得一批原创研究成果，以期在丰富和完善多分类器系统理论与方法的工作中有所突破，并为新成果在复杂模式分类问题中得以成功应用，提供坚实的理论基础与关键技术储备。