基于增强AlexNet的音乐流派识别研究

PDF下载HTML

Title:

Music genre recognition research based on enhanced AlexNet

作者:

刘万军, 孟仁杰, 曲海成, 刘腊梅

辽宁工程技术大学 软件学院，辽宁 葫芦岛 125105

Author(s):

LIU Wanjun, MENG Renjie, QU Haicheng, LIU Lamei

College of Software, Liaoning Technical University, Huludao 125105, China

关键词:

音乐流派识别; 深度卷积神经网络; 机器学习; 深度学习; AlexNet; 音频特征提取; 音乐特征识别

Keywords:

music genres recognition; deep convolutional neural network; machine learning; deep learning; AlexNet; audio feature extraction; audio feature extraction

分类号:

TP181

DOI:

10.11992/tis.201909032

摘要:

针对机器学习模型对音乐流派特征识别能力较弱的问题，提出了一种基于深度卷积神经网络的音乐流派识别(DCNN-MGR)模型。该模型首先通过快速傅里叶变换提取音频信息，生成可以输入DCNN的频谱并切割生成频谱切片。然后通过融合带泄露整流(Leaky ReLU)函数、双曲正切(Tanh)函数和Softplus分类器对AlexNet进行增强。其次将生成的频谱切片输入增强的AlexNet进行多批次的训练与验证，提取并学习音乐特征，得到可以有效分辨音乐特征的网络模型。最后使用输出模型进行音乐流派识别测试。实验结果表明，增强的AlexNet在音乐特征识别准确率和网络收敛效果上明显优于AlexNet及其他常用的DCNN、DCNN-MGR模型在音乐流派识别准确率上比其他机器学习模型提升了4%~20%。

Abstract:

To solve the problem that machine learning model has weak ability to identify music genre features, a music genre recognition model based on deep convolutional neural network (DCNN-MGR) is proposed in this paper. At first, the model extracts audio information through Fast Fourier Transformation, generating spectrums that can be input to the DCNN and slicing the generated spectrums. Then AlexNet is enhanced by fusion of Leaky ReLU function, Tanh function and Softplus classifier. The generated spectrum slices are input into the enhanced AlexNet for multi-batch training and verification. Music features are extracted and learned, and a network model that can effectively distinguish music features is obtained. At last, the output model is applied to music genre recognition and test. The experimental results show that the enhanced AlexNet is superior to AlexNet and other commonly used DCNN in terms of accuracy of music feature recognition and network convergence effect. The DCNN-MGR model is 4%~20% higher than other machine learning models in music genre recognition accuracy.