基于小波系数聚类的特征提取分类方法.PDFVIP

第 38 卷 第 8 期 2001 年 8 月 计 算 机 研 究 与 发 展 JOU RNAL O F COM PU T ER R ESEA RCH D EV ELO PM EN T V o l138,N o18 A ug. 2001 原稿收到日期: 2000207205; 修改稿收到日期: 2001203219 基于小波系数聚类的特征提取分类方法 周维忠①　冯心海①　孙国基② ① (佛山科学技术学院电子工程系　佛山　528000) ② (西安交通大学系统工程研究所　西安　710049) (fszw z@ 21cn. com ) 摘　要　神经网络是一种普遍采用的模式分类方法, 当对样本的抽样数目较大时, 神经网络结构复杂, 训练时间激 增, 分类性能下降, 针对这一问题, 提出一种基于快速小波变换特征提取的分类方法. 首先对小波系数矩阵的每行 进行聚类, 表达重要频率范围内小波系数矩阵的行有较多的聚类数, 从而大大减少了神经网络的输入数, 而同时保 留了有用的信息. 特征提取后, 采用小波系数的能量值作为特征量, 应用径向基函数网络识别肺发出的各种不同的 声音, 实验证明: 该方法有较高的识别率. 关键词　特征提取, 快速小波变换, 径向基函数神经网络 中图法分类号　T P391; T P18 CLASSIF ICAT ION BASED ON FEATURE EXTRACT ION FROM CL USTER OF W AVEL ET COEFF IC IENTS ZHOU W ei2Zhong①, FEN G X in2H ai①, and SUN Guo2J i② ① (D ep artm en t of E lectron ic E ng ineering , F oshan Institu te of S cience and T echnology , F oshan 528000) ② (S y stem E ng ineering Institu te, X i’an J iaotong U niversity , X i’an 710049) Abstract　N eu ra l netw o rk is popu larly u sed fo r pa t tern recogn it ion. T he tra in ing t im e of neu ra l netw o rk increases rap id ly w ith increasing large num ber of sam p les. It leads to a deterio ra t ion in perfo rm ance of neu ra l netw o rk. A classif ica t ion app roach based on fast w avelet t ran sfo rm fo r fea tu re ex tract ion is p resen ted. T he m ethod divides the m atrix of com pu ted w avelet coefficien ts in to clu sters in every row s. T he row s tha t rep resen t im po rtan t frequency range have a larger num bers of clu sters than row s tha t rep resen t less im po rtan t frequency ranges. T he inpu t num bers of neu ra l netw o rk are decreased, w h ile reta in ing m uch info rm at ion abou t classif ied signa l. A fter fea tu re ex tract ion, energy values of w avelet coefficien ts are cho sen as signa l fea tu res. A rad ia l basic funct ion neu ra l netw o rk is developed fo r classif ica t ion of d ifferen t lung sound signa ls. T he effect iveness of th is new m ethod has been verif ied on experim en ts abou t recogn izing lung sounds. Key words