基于稀疏表示和深度神经网络模型的西洋乐器自动分类方法

发布时间：2018-03-14 23:05

本文选题：字典库　切入点：稀疏重构　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着现代社会科技的快速发展,数字音乐的数量也以海量的方式增长,为了方便用户音乐检索,对音乐进行有效合理的分类十分重要。但音乐基数多,每天新的单曲数又不断增加,采取人工的方式对音乐进行分类不切实际。因此,本文采用深度神经网络模型对音乐提取特征实现自动分类。与此同时,由于音乐信号具备稀疏性的特点,本文创新型地提出将稀疏特征与深度神经网络模型相结合的方法,实现音乐信号的自动分类。音乐信号分类标准众多,本文旨在将提出的新方法应用在西洋乐器的自动分类上。本文首先介绍了音乐信号的常见特征,包括基本的物理学特征和人耳感知的心理学角度的特征,与此同时还介绍了人耳听觉系统的特性。接着介绍了音乐自动分类的理论,包含对音乐的特征提取,并着重介绍了本文所采用的深度神经网络模型方面的相关概念。然后研究了音乐信号的稀疏特征提取过程,创新性地引入新型的稀疏表示字典库,该字典库是基于不同乐器发出音色频率间的差异。本文通过MATLAB仿真实验,对比了基于传统字典库和基于本文构建字典库下的稀疏重构结果,证明了基于本文构建字典库下进行稀疏重构效果优于传统字典库。最后本文研究了基于稀疏表示和深度神经网络模型结合的西洋乐器自动分类情况,传统的深度神经网络模型输入端为音频信号的梅尔频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC),本文创新性地使用音频信号的稀疏特征作为深度神经网络模型输入端,在此基础上训练深度神经网络模型参数,实现西洋乐器的自动分类。本文采用Python脚本语言,通过实验证明了基于稀疏特征和深度神经网络模型相结合的方法,实现西洋乐器自动分类,准确率可达82%。
[Abstract]:With the rapid development of modern social science and technology, the number of digital music is also increasing in a massive way. In order to facilitate users' music retrieval, it is very important to classify music effectively and reasonably. The number of new single songs is increasing every day, so it is impractical to classify music manually. Therefore, the depth neural network model is used to realize the automatic classification of music features in this paper. Due to the sparsity of music signals, this paper proposes an innovative method to combine sparse features with depth neural network model to realize the automatic classification of music signals. The purpose of this paper is to apply the proposed new method to the automatic classification of western musical instruments. This paper first introduces the common features of musical signals, including the basic physical features and the psychological features of human ear perception. At the same time, it also introduces the characteristics of human ear and hearing system, and then introduces the theory of automatic music classification, including the feature extraction of music. Then the sparse feature extraction process of music signal is studied, and a new sparse representation dictionary is introduced. The dictionary is based on the difference of timbre frequency between different musical instruments. This paper compares the sparse reconstruction results based on the traditional dictionary library and the dictionary database based on this paper through MATLAB simulation experiments. It is proved that the effect of sparse reconstruction based on this paper is better than that of traditional dictionary library. Finally, the paper studies the automatic classification of western musical instruments based on sparse representation and depth neural network model. The traditional depth neural network model input end is Mel Frequency Cepstrum efficient Mel frequency cepstrum. In this paper, the sparse feature of audio signal is innovatively used as the input of depth neural network model. On this basis, the parameters of the deep neural network model are trained to realize the automatic classification of western musical instruments. In this paper, the method based on sparse feature and depth neural network model is proved by using Python script language. The automatic classification of western musical instruments is realized, and the accuracy rate is up to 82.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TN912.3

【相似文献】