一种改进的类别依赖型特征选择技术

英文篇名：An Improved Feature Extraction Technology Tied to Different Classes
中文刊名：计算机与数子工程
英文刊名：Computer & Digital Engineering
作者：贾沛 ; 桑农 ; 唐红卫
英文作者：Jia Pei Sang Nong Tang Hongwei(Institute for Pattern Recognition and Artificial Intelligence ; HUST ; State Education Commission Key Laboratory for Image Processing and Intelligent Control ; Wuhan 430074)
中文关键词：特征选择 ; 分类识别 ; BP神经网络
英文关键词：Feature Selection ; Classification Recognition ; BP Neural Network
出版日期：2003-12-20
机构：华中科技大学图像识别与人工智能研究所图像信息处理与智能控制教育部重点实验室;
年：2003
期：06
出版单位：计算机与数子工程

摘要

特征选择是模式识别领域的重要环节。本文提出了一种改进的类别依赖型特征选择技术,通过事先选定参数m,自动选择出在文中准则下各个类各自最能区别于其他类的k_x维特征。将各类所选择特征的并集作为BP神经网络的输入结点,进而用待识别样本进行分类识别。实验结果说明,该类特征选择方法能够选择出对每一类而言最能区分于其他类的特征。与选择所有特征相比,用该法进行特征选择以后的BP神经网络分类识别有着较高的正确识别率。此外,本文还利用相关分析方法在预处理过程中剔除了线性相关的冗余特征。
Feature selection is an important part of pattern recognition.This paper provides an improved technology of feature selection tied to different classes.Through adjusting m,k_x features of a certain class x can be chosen out automatically.We combine all the chosen features tied to different classes and take them as the inputs of BP neural network for classification.The result of the experiment shows that this method can choose out the most representative and the most separable features of different classes.Compared with others the correct classification ratio is high as well.Besides,all the linear correlated features have been removed during preprocessing.

引文

[1] 李金宗模式识别导论,高等教育出版社,1994
    [2] 边肇祺张学工等模式识别,清华大学出版社,2000
    [3] Orly Yadid-pecht,Moshe Gur,Efficient biologically-based pattern-recognizing networks,Neural Networks,Vol.9,No.6,pp.1060-1070,1996
    [4] 吉小军李世中李霆,相关分析在特征选择中的应用,华北工学院测试技术学报,Vol.15,No.1,2001
    [5] 王裕玲李磊常希光,地震模式识别特征选择的方法及实现,石油地球物理勘探,第30卷增刊,1995,12