基于并行胶囊网络的声学场景分类  

Acoustic scene classification based on parallel capsule network

作　　者：杨立东[1] 赵飞焱 YANG Lidong;ZHAO Feiyan(School of Information Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,China)

机构地区：[1]内蒙古科技大学信息工程学院,内蒙古包头014010

出　　处：《传感器与微系统》2023年第12期155-159,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(62161040);内蒙古自然科学基金资助项目(2021MS06030);内蒙古科技计划资助项目(2021GG0023)。

摘　　要：为解决卷积神经网络(CNN)忽略音频特征之间的空间关系、丢失姿态特征和时序性特征的问题,提出了基于并行胶囊网络的声学场景分类模型,选用胶囊网络和双向门控循环单元弥补CNN的缺陷。首先,该模型通过提取音频对数梅尔能量谱特征;然后,结合各模块优点对音频特征处理;最后,根据场景特征完成分类。通过在“国际声学场景和事件检测及分类(DCASE)挑战赛2019”挑战任务1数据集下进行实验,在开发集和验证集上分别获得了71.1%和70.2%的准确率,优于基线系统的准确率,证明了该网络模型适用于声学场景分类任务。In order to solve the problem that the spatial relationship between audio features,lost of posture features and timing features are ignored by the convolutional neural network(CNN),an acoustic scene classification model based on parallel capsule network is proposed.Capsule network and bidirectional gating recurrent unit are selected to compensate for the defects of CNN.Firstly,the audio log-Mel energy spectrum features are extracted by model.Then,the audio features are processed by combining the advantages of each module.Finally,the classification task is completed according to the scene features.Experiments are carried out on the dataset of Detection and Classification of Acoustic Scenes and Events(DCASE)2019 Challenge Task 1.The accuracy rates of 71.1%and 70.2%are obtained on the development set and validation set,respectively,which are better than the accuracy of the baseline system.It is proved that the network model is suitable for the acoustic scene classification task.

关 键 词：声学场景分类 胶囊网络 双向门控循环单元 并行神经网络 动态路由机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP212[自动化与计算机技术—计算机科学与技术]