基于小波分解和Mel频率的儿童咳嗽干湿性自动分类

AUTOMATIC CLASSIFICATION OF CHILDHOOD DRY AND WET COUGH BASED ON WAVELET PACKET DECOMPOSITION AND MEL-FREQUENCY

咳嗽的干湿性是儿科呼吸道疾病诊断的重要依据。咳嗽干湿性自动分类一般以咳嗽音MFCC作为主要特征向量,但Mel频率滤波器组低频密集高频稀疏的分布特性,使其未充分利用不同频段在反映干湿性能力上的差异。小波包分解可以获得不同频段的特征,对小波能量进行Mel频率刻度的非线性伸缩,以弱化低频特征、强化高频特征,从而凸显两类咳嗽音信号在各频段上的差异性。实验结果表明,该特征向量能更有效地区分两类样本数据,干湿性总体分类准确率达89. 6%。

The classification of dry and wet cough plays an important role in the childhood diagnosis of pediatric respiratory diseases. The traditional classification of wet and dry cough usually extracts the MFCC as the main feature vector, but the distribution characteristics of Mel frequency filters are dense in low-frequency and sparse in high-frequency, so it not fully utilize the difference between various frequency bands in reflecting dry and wet. Wavelet packet decomposition can get the characteristics of various frequency bands. The nonlinear expansion of the Mel frequency scale of the wavelet energy was carried out to weaken low-frequency characteristics and enhance high-frequency characteristics, thus the difference of the two kinds of cough sound signals in each frequency band was highlighted. The experimental results show that the feature vector can effectively classify two types of sample data, and the overall accuracy of dry and wet classification is 89.6%.