为解决传统城市噪音分类中数据过少而导致模型泛化效果不好,鲁棒性过高,同时传统的噪音特征不能解决关键数据丢失问题导致模型准确率下降。本文提出了一种基于MFCC + GFCC混合特征和噪音语谱图特征的双路卷积模型。该模型首先对噪音数...为解决传统城市噪音分类中数据过少而导致模型泛化效果不好,鲁棒性过高,同时传统的噪音特征不能解决关键数据丢失问题导致模型准确率下降。本文提出了一种基于MFCC + GFCC混合特征和噪音语谱图特征的双路卷积模型。该模型首先对噪音数据进行MFCC,GFCC和语谱图变化,提取特征数据,将MFCC和GFCC数据分别进行卷积压缩处理,并在混合后进行分类。对于噪音的语谱图特征进行卷积后,使用注意力机制模块对其各个通道进行权重标记后进行分类,将两路的分类结果进行贝叶斯数值融合,从而实现对城市噪音的正确分类。实验结果表明,识别的准确率比传统模型网络在大数据样本的数据集下有了8%左右以上的提升。In order to solve the problem of poor generalization effect and high robustness of the model due to too little data in the traditional urban noise classification, the accuracy of the model decreases due to the fact that the traditional noise features cannot solve the problem of key data loss. In this paper, a two-way convolution model based on MFCC + GFCC hybrid features and noise spectral features is proposed. Firstly, the noise data is changed by MFCC, GFCC and spectrogram, the feature data is extracted, the MFCC and GFCC data are convoluted and compressed respectively, and the classification is carried out after mixing. After convoluting the spectral features of noise, the attention mechanism module is used to classify each channel by weighting labeling, and the classification results of the two channels are fused with Bayesian numerical values, so as to achieve the correct classification of urban noise. Experimental results show that the accuracy of recognition is improved by more than 8% compared with the traditional model network under the dataset of big data samples.展开更多
基金本研究得到以下两个项目支持:Yunfei Du,School of Basic Education,Beijing Institute of Graphic Communication,Beijing 102600,China,项目:the Project of Beijing Municipal Commission of Education (KM 202110015001)北京印刷学院重点教学改革项目——工程认证背景下的工科数学教学改革对大学生创新思维与创业能力培养的研究与实践。
文摘为解决传统城市噪音分类中数据过少而导致模型泛化效果不好,鲁棒性过高,同时传统的噪音特征不能解决关键数据丢失问题导致模型准确率下降。本文提出了一种基于MFCC + GFCC混合特征和噪音语谱图特征的双路卷积模型。该模型首先对噪音数据进行MFCC,GFCC和语谱图变化,提取特征数据,将MFCC和GFCC数据分别进行卷积压缩处理,并在混合后进行分类。对于噪音的语谱图特征进行卷积后,使用注意力机制模块对其各个通道进行权重标记后进行分类,将两路的分类结果进行贝叶斯数值融合,从而实现对城市噪音的正确分类。实验结果表明,识别的准确率比传统模型网络在大数据样本的数据集下有了8%左右以上的提升。In order to solve the problem of poor generalization effect and high robustness of the model due to too little data in the traditional urban noise classification, the accuracy of the model decreases due to the fact that the traditional noise features cannot solve the problem of key data loss. In this paper, a two-way convolution model based on MFCC + GFCC hybrid features and noise spectral features is proposed. Firstly, the noise data is changed by MFCC, GFCC and spectrogram, the feature data is extracted, the MFCC and GFCC data are convoluted and compressed respectively, and the classification is carried out after mixing. After convoluting the spectral features of noise, the attention mechanism module is used to classify each channel by weighting labeling, and the classification results of the two channels are fused with Bayesian numerical values, so as to achieve the correct classification of urban noise. Experimental results show that the accuracy of recognition is improved by more than 8% compared with the traditional model network under the dataset of big data samples.
