大间距无相关边界Fisher鉴别分析  

Analysis of Maximum Marginal Uncorrelated Marginal Fisher

作　　者：梁兴柱[1] 林玉娥[1] 李敬兆[1] 林玉荣[2] 

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]哈尔滨工业大学航天学院,黑龙江哈尔滨150001

出　　处：《山东科技大学学报（自然科学版）》2013年第1期84-87,共4页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：国家自然科学基金项目(60975009);安徽省自然科学基金项目(1208085QF123);安徽省高等学校自然科学基金项目(KJ2012Z084);安徽理工大学青年教师科学研究基金项目(2012QNZ10)

摘　　要：针对边界Fisher鉴别分析方法存在的小样本问题以及所求出的鉴别矢量集缺少约束限制的缺陷,提出了一种大间距无相关边界Fisher鉴别分析方法。该方法采用最大化描述样本数据可分性和紧致性的矩阵之差作为目标函数,避免了边界Fisher鉴别分析的小样本问题;对于无相关鉴别矢量集的求解,给出了先构造无相关空间,再进行特征值分解的求解策略。仿真结果表明,该方法在识别性能上优于已有的边界Fisher鉴别分析及其改进方法,且避免了使用繁琐的迭代方法求解不相关鉴别矢量集,具有一定的实用价值。Aiming at the shortcomings of the marginal Fisher analysis, such as small sample size problems and obtained discriminant vectors lack of constraints, a maximum marginal uncorrelated marginal Fisher analysis was proposed. By adopting the maximum margin between matrices which characterize the divisibility and compactness of data as the object function, the proposed method avoids the small sample size problems in the marginal Fisher analysis. As to the computation of the uncorrelated discriminant vectors,the uncorrelated space is computed firstly and then the standard eigenvalue problem is solved. Simulation results show that the new method, which is superior to the existing marginal fisher analysis and its improved methods in the recognition performance, also avoids using the iterate method for the uncorrelated discriminant vectors and has some practical value.

关 键 词：大间距 无相关 边界Fisher鉴别分析 目标函数 小样本问题 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]