基于统计学习理论的快速算法及其应用研究

项目名称： 基于统计学习理论的快速算法及其应用研究

项目编号： No.10801004

项目类型： 青年科学基金项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 韩敏

作者单位： 北京工业大学

项目金额： 17万元

中文摘要： 本项目研究了基于统计学习理论的支持向量机、Neyman-Pearson 学习等的快速收敛算法及其推广,并将其应用于模式识别、计算机视觉等领域中。在一定的条件下使构造的算法具有指数型衰减的逼近误差和取样误差，分析算法的相容性条件，以及收敛速度、计算复杂度等问题。构造了基于Rademacher 复杂度的Neyman-Pearson 经验风险算法，研究其快速收敛性。并且构造了半监督的基于Neyman-Pearson学习支持向量机算法，该算法具有快速收敛、损失敏感的特点。基于统计学习理论的快速算法在两类模式识别问题中已经应用比较广泛，在复杂的计算机视觉、人脸识别等领域中的应用正在引起人们的关注，本项目还研究了统计学习理论的快速算法在计算机视觉领域中的应用，并将前面最新的研究成果编程，通过真实数据加以例证。

中文关键词： 统计学习理论；快速算法；支持向量机；Neyman-Pearson 学习；计算机视觉

英文摘要： This project investigates some fast learning algorithms based on statistical learning theory such as support vector machine，Neyman-Pearson learning algorithms. The approximations for Bayes function and applications to pattern recognition, computer vision of all this algorithms are also discussed. Under certain conditions our algorithms convergence exponentially both in the approximation error and sampling error. We analysis the conditions for fast algorithms. We construct empirical risk minimization algorithms based on Rademacher complexity Neyman-Pearson learning. The fast convergence boundary is rereceived. We construct a semi-supervised sopport vector machine algorithm based on Neyman-Pearson learning. The algorithm has fast converge rate and be loss-sensitive. Statistical learning theory algorithms has been used widely in the two-class pattern recognition problems, but the applications in computer vision, face recognition and other areas are in progress.The applications of fast learning algorithms based on statistical learning theory to the field of computer vision are just the study of this project.We program our research results and experiment on real data.

英文关键词： statistical learning theory;fast algorithms;support vector machine;Neyman-Pearson learing;computer vision