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
Neural network is popularly used for pattern recognition. The training time of neural network increases rapidly with increasing large number of samples. It leads to a deterioration in performance of neural network. A classification approach based on fast wavelet transform for feature extraction is presented. The method divides the matrix of computed wavelet coefficients into clusters in every rows. The rows that represent important frequency range have a larger numbers of clusters than rows that represent less important frequency ranges. The input numbers of neural network are decreased, while retaining much information about classified signal. After feature extraction, energy values of wavelet coefficients are chosen as signal features. A radial basic function neural network is developed for classification of different lung sound signals. The effectiveness of this new method has been verified on experiments about recognizing lung sounds.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2001年第8期982-987,共6页
Journal of Computer Research and Development