基于特征融合的生猪音频识别  

Pig Audio Recognition Based on Feature Fusion

作　　者：熊婷荣 陈威 王宁 汤先美 辜丽川[1] 焦俊[1] XIONG Tingrong;CHEN Wei;WANG Ning;TANG Xianmei;GU Lichuan;JIAO Jun(College of Information and Computer,Anhui Agricultural University,Hefei 230036,China)

机构地区：[1]安徽农业大学信息与计算机学院,合肥230036

出　　处：《合肥学院学报（综合版）》2024年第2期100-107,共8页Journal of Hefei University：Comprehensive ED

基　　金：安徽省科技重大专项项目“大数据环境下的生猪健康养殖与疫病防控预警关键技术研究”(201903a06020009,202103b06020013);2021年度校研究生教育教学质量工程项目(2021yjsjd03)。

摘　　要：针对在生猪音频识别中单一特征参数无法充分地表征猪声信息的问题,提出了基于梅尔倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)与伽马通倒谱系数(Gammatone Frequency Cepstral Coefficient,GF-CC)的生猪音频融合特征生成方法。首先,以5种猪声为研究对象,利用功率谱减法和双门限端点检测法对猪声样本进行预处理。其次,提取MFCC、GFCC和它们的一阶差分参数,将MFCC+∆MFCC、GFCC+∆GFCC直接叠加得到高维的融合特征,为了降低高维特征的冗余度,利用增减分量法对其进行降维,最后将降维后的融合特征输入至Bi-LSTM网络模型进行分类识别。实验结果表明,相对于传统的单一特征MFCC、GFCC在识别率上分别提升了14.33%和18.63%,且在不同噪声环境下,融合特征具有比其他特征更好的鲁棒性和识别性能。Aiming at the problem that single feature parameter cannot fully represent the pig sound in-formation in pig audio recognition,this paper proposed a pig audio fusion feature generation method based on Mel-cepstral coefficient and Gammatone cepstral coefficient.Firstly,five pig sounds were taken as research objects,and the pig sound samples were preprocessed by power spectral subtraction and double threshold endpoint detection.Secondly,MFCC,GFCC and their first-order difference pa-rameters were extracted,and the high-dimensional fusion features were obtained by directly stacking MFCC+∆MFCC and GFCC+∆GFCC.In order to reduce the redundancy of high-dimensional features,the dimension of the fusion features was reduced by using the increasing and decreasing component method.Finally,the fusion features after dimension reduction were input into the Bi-LSTM network model for classification and recognition.Experimental results show that compared with the traditional single feature MFCC and GFCC,the recognition rate is increased by 14.33%and 18.63%respectively,and the fusion feature has better robustness and recognition performance than other features in differ-ent noise environments.

关 键 词：生猪音频识别 特征融合 MFCC Gammatone滤波器 

分 类 号：S828[农业科学—畜牧学] TN912.34[农业科学—畜牧兽医]