基于特征级联的肺炎咳嗽声识别

Pneumonia cough sound recognition based on feature cascade

针对识别肺炎咳嗽声方面现有特征组合方法单一的问题,提出了实现级联浅层和深层特征的SELeNet-5网络模型。首先对咳嗽声使用6层小波包分解得到相对小波包能量作为浅层特征,同时,对咳嗽声预处理生成语谱图,使用改进的LeNet-5网络作为特征提取模型,从语谱图提取出深层特征;其次通过特征级联网络层结合浅层和深层两种不同层次的特征,形成新的特征向量;最终输入到注意力机制网络模块并通过全连接网络层输出肺炎咳嗽声的识别结果。对比实验结果表明,使用特征级联方法的SELeNet-5网络模型方法得到了79.81%的准确率,相比使用单一浅层特征准确率提高了6.81%,与使用单一深层特征相比准确率提高了2.92%。实验结果表明级联特征在肺炎咳嗽声识别上比单一的浅层或深层特征具有更好的效果,有效提高了肺炎咳嗽声识别准确率。

The existing feature combination methods for identifying pneumonia cough sound are single,so a SELeNet-5 network model that realizes cascading shallow and deep feature is proposed.The cough sound is subjected to 6-layer wavelet packet decomposition to obtain the relative wavelet packet energy which is taken as the shallow feature.The cough sound is preprocessed to produce a speech spectrogram.The improved LeNet-5 network is used as the feature extraction model to extract the deep features from the speech spectrogram.A new feature vector is formed by combining the shallow layer and deep layer features with the feature cascade network layer.It is input into the attention mechanism network module,and the recognition result of pneumonia cough sound is output by the fully-connected network layer.The results of comparative experiment show that the SELeNet-5 network model method based on the feature cascade reaches an accuracy of 79.81%,which is 6.81% higher than the method based on single shallow feature,and 2.92% higher than the method based on single deep feature.The experimental results show that the cascade feature has a better effect on the recognition of pneumonia cough sound than single shallow feature or deep feature,which effectively improves the recognition accuracy of pneumonia cough sound.