N-DenseNet的城市声音事件分类模型

doi:10.19665/j.issn1001-2400.2019.06.002

摘要/Abstract

摘要：

针对城市声音事件分类领域中现有模型分类准确率不高、泛化能力不强的问题,提出了一种N阶密集卷积神经网络的城市声音事件分类模型。首先,介绍了密集卷积神经网络的结构;其次,基于N阶马尔可夫模型将密集连接改进为N阶有关连接;然后,结合两者提出了一种更适合音频分类的模型——N阶密集卷积神经网络。该模型在避免梯度消失的前提下, 有针对性、规律性减少了特征图层之间的连接,更高效地融合了前N特征图层的信息,使得模型的收敛速度更快;最后,为了验证该模型,采用N阶密集卷积神经网络的一阶、二阶子模型,基于UrbanSound8K和Dcase2016数据集开展了城市声音事件分类研究。研究结果表明,其模型准确率分别为83.63%、81.03%,验证了该模型具有良好的分类准确率和泛化能力。

关键词: 声音事件分类, 密集卷积神经网络, N阶马尔可夫模型, N阶密集卷积神经网络

Abstract:

An urban sound event classification model based on the N-order Dense Convolutional Network (abbreviated to N-DenseNet) is proposed for the problems of insufficient classification accuracy and generalization ability of existing models. First, the network structure of the DenseNet is briefly introduced. Then, dense connection in the DenseNet is improved by N-order state-dependent connection based on the N-order Markov model. Furthermore, combining advantages of both the DenseNet and N-order Markov, a novel network architecture, i.e., the N-DenseNet, is proposed in this paper. Theoretically, the N-DenseNet satisfying the premise of alleviating vanishing-gradient, can not only produce efficient integration of feature information from the layers, but also accelerate the convergence speed. Finally, in order to validate advantages of the new model, 1-DenseNet and 2-DenseNet are respectively exploited in the urban sound event classification based on the UrbanSound8K and Dcase2016 dataset. Experimental results show that the accuracy of the two above-mentioned models is respectively 83.63% and 81.03%, which also demonstrates a higher classification accuracy and a better generalization performance of the N-DenseNet.

Key words: sound event classification, dense convolutional network, N-order Markov model, N-order dense convolutional

中图分类号:

TP391.42

曹毅,黄子龙,张威,刘晨,李巍. N-DenseNet的城市声音事件分类模型[J]. 西安电子科技大学学报, 2019, 46(6): 9-16.

CAO Yi,HUANG Zilong,ZHANG Wei,LIU Chen,LI Wei. Urban sound event classification with the N-order dense convolutional network[J]. Journal of Xidian University, 2019, 46(6): 9-16.

图/表 11

图1

图2

图3

表1

表2

图4

图5

表3

表4

图6

图7

参考文献 19

[1]	BOGDANOV D, WACK N, GOMEZ E , et al. An Open-source Library for Sound and Music Analysis[C]//Proceedings of the 2013 ACM Multimedia Conference. New York: ACM, 2013: 855-858.
[2]	王瑞, 王康晏, 冯玉田 , 等. 复杂场景下声频传感器网络核稀疏表示车辆识别[J]. 西安电子科技大学学报, 2015,42(4):114-120.
	WANG Rui, WANG Kangyan, FENG Yutian , et al. Vehicle Recognition Using Acoustic Sensor Networks in Complex Scenes via Kernel Sparse Representation[J]. Journal of Xidian University, 2015,42(4):114-120.
[3]	DUBOIS D, GUASTAVINO C, MAFFIOLO V , et al. A Cognitive Approach to Soundscape Research[J]. Journal of the Acoustical Society of America, 2004,115(5):2495. doi: 10.1080/00140139.2013.835873 pmid: 24073684
[4]	HUSSEIN R, SHABAN K B, EL-HAG A H . Robust Feature Extraction and Classification of Acoustic Partial Discharge Signals Corrupted with Noise[J]. IEEE Transactions on Instrumentation and Measurement, 2017,66(3):405-413. doi: 10.1109/TIM.2016.2639678
[5]	郭晶晶, 马建峰 . 面向虚拟社区物联网的信任推荐算法[J]. 西安电子科技大学学报, 2015,42(2):52-57. doi: 10.3969/j.issn.1001-2400.2015.02.009
	GUO Jingjing, MA Jianfeng . Trust Recommendation Algorithm for the Virtual Community Based Internet of Things(IoT)[J]. Journal of Xidian University, 2015,42(2):52-57. doi: 10.3969/j.issn.1001-2400.2015.02.009
[6]	王静远, 李超, 熊璋 , 等. 以数据为中心的智慧城市研究综述[J]. 计算机研究与发展, 2014,51(2):239-259.
	WANG Jingyuan, LI Chao, XIONG Zhang , et al. Survey of Data-centric Smart City[J]. Journal of Computer Research and Development, 2014,51(2):239-259.
[7]	SALAMON J, JACOBY C, BELLO J P . A Dataset and Taxonomy for Urban Sound Research[C]//Proceedings of the 2014 ACM Conference on Multimedia. New York: ACM, 2014: 1041-1044.
[8]	YE J X, KOBAYASHI T, MURAKAWA M . Urban Sound Event Classification Based on Local & Global Features Aggregation[J]. Applied Acoustics, 2017,117:246-256. doi: 10.1016/j.apacoust.2016.08.002
[9]	KONS Z, TOLEDO-RONEN O . Audio Event Classification Using Deep Neural Networks[C]//Proceedings of the 2013 Annual Conference of the International Speech Communication Association. Baixas: International Speech Communication Association, 2013: 1482-1486.
[10]	LIM M, LEE D, PARK H , et al. Convolutional Neural Network Based Audio Event Classification[J]. KSII Transactions on Internet and Information Systems, 2018,12(6):2748-2760. doi: 10.1371/journal.pone.0214168 pmid: 31647815
[11]	CHEN Y, GUO Q, LIANG X Y , et al. Environmental Sound Classification with Dilated Convolutions[J]. Applied Acoustics , 2019,148:123-132. doi: 10.1016/j.apacoust.2018.12.019
[12]	ZHANG X H, ZOU Y X, SHI W . Dilated Convolution Neural Network with Leaky ReLU for Environmental Sound Classification[C]//Proceedings of the 2017 International Conference on Digital Signal Processing. Piscataway: IEEE, 2017: 8096153.
[13]	HUANG G, LIU Z, VAN DER MAATEN L , et al. Densely Connected Convolutional Networks[C]//Proceedings of the 2017 30th IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 2261-2269.
[14]	吴仁彪, 赵婷, 屈景怡 . 基于深度SE-DenseNet的航班延误预测模型[J]. 电子与信息学报, 2019,41(6):1510-1517. doi: 10.1021/acsnano.9b07956 pmid: 31869204
	WU Renbiao, ZHAO Ting, QU Jingyi . Flight Delay Prediction Model Based on Deep SE-DenseNet[J]. Journal of Electronics & Information Technology, 2019,41(6):1510-1517. doi: 10.1021/acsnano.9b07956 pmid: 31869204
[15]	JOAKIM M, JOAKIM W . An N-state Markov-chain Mixture Distribution Model of the Clear-sky Index[J]. Solar Energy, 2018,173(1):487-495. doi: 10.1016/j.solener.2018.07.056
[16]	曹毅, 黄子龙, 张威 , 等. 基于N-DenseNet和高维mfcc特征的城市声音事件分类方法: CN109949824A [P]. 2019-06-28.
[17]	MESAROS A, HEITTOLA T, VIRTANEN T . TUT Database for Acoustic Scene Classification and Sound Event Detection[C]//Proceedings of the 2016 European Signal Processing Conference. Poland: European Signal Processing Conference, 2016: 1128-1132.
[18]	李东, 张雪英, 段淑斐 , 等. 结合语音融合特征和随机森林的构音障碍识别[J]. 西安电子科技大学学报, 2018,45(3):149-155.
	LI Dong, ZHANG Xueying, DUAN Shufei , et al. Dysarthria Recognition Combining Speech Fusion Feature and Random Forest[J]. Journal of Xidian University, 2018,45(3):149-155.
[19]	仲伟峰, 方祥, 范存航 , 等. 深浅层特征及模型融合的说话人识别[J]. 声学学报, 2018,43(2):60-70.
	ZHONG Weifeng, FANG Xiang, FAN Cunhang , et al. Fusion of Deep Shallow Feature and Models for Speaker Recognition[J]. Acta Acustica, 2018,43(2):60-70.

特征图层数L	Block结构	有关连接数	模型有关连接数	相对减少率/%
5	Dense Block	10	10×m
5	1-DenseBlock	4	4×m	60
5	2-DenseBlock	6	6×m	40
L	Dense Block	L(L-1)/2	mL(L-1)/2
L	N-DenseBlock	N(L-N)	mN(L-N)	33(Min)

模型	有关连接	总参数量	模型大小/MB	训练时间/min	趋于收敛/epoch
DenseNet-16	9	1 655 730	19.62	77	45
1-DenseNet-16	6	1 652 946	19.59	72	30
2-DenseNet-16	6	1 572 306	18.64	63	30

网络模型	特征	准确率/%
DenseNet-16	128mfcc	80.16
1-DenseNet-16	128mfcc	82.03
2-DenseNet-16	128mfcc	81.48
2-DenseNet-22	128mfcc	81.96
DenseNet-16	128mfcc+23mfcc_d+23mfcc_d_d	81.00
1-DenseNet-16	128mfcc+23mfcc_d+23mfcc_d_d	83.63
2-DenseNet-16	128mfcc+23mfcc_d+23mfcc_d_d	82.17
2-DenseNet-22	128mfcc+23mfcc_d+23mfcc_d_d	83.27
SVM[7]	25mfcc	71.00
DCNN[8]	Local global features aggregation	77.36
DilatedCNN[11]	64n_mels spectrograms	78.00
DNN[9]	80FBANK	79.23
CNN[10]	40FBANK	81.50
D-CNN-ESC[12]	60mfcc+60mfcc_d	81.90

Model	UrbanSound8K	Dcase2016
Baseline	71.00	77.20
DenseNet-16	81.00	78.28
1-DenseNet-16	83.63	80.00
2-DenseNet-16	82.17	81.03
2-DenseNet-22	83.27	80.28