基于SE-ResNet18的德州驴声音分类研究  

Classification Research of Dezhou Donkey Sound Based on SE-ResNet18

作　　者：徐宗鹏 沈延锋 张亚群 潘雨龙 王广超 于家峰 XU Zongpeng;SHEN Yanfeng;ZHANG Yaqun;PAN Yulong;WANG guangchao;YU Jiafeng(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin Jilin 132022,China;Shandong Key Laboratory of Biophysics,Institute of Biophysics,Dezhou University,Dezhou Shandong 253023,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林吉林132022 [2]德州学院生物物理研究院山东省生物物理重点实验室,山东德州253023

出　　处：《德州学院学报》2024年第6期35-40,共6页Journal of Dezhou University

基　　金：山东省重点研发计划(农业良种工程)(2023LZGCQY020)。

摘　　要：声音与畜牧健康养殖具有密切关联,但由于数据采集难度大等原因,对动物声音的研究还具有较大挑战。以我国特有的德州驴为研究对象,结合通道注意力机制的SE-ResNet18模型提出了一种德州驴声音分类算法。首先,采用谱减法和单参数双门限端点检测法对采集到的三类状态下驴的声音信号进行预处理,并提取13维梅尔倒谱系数为特征。其次,将SE模块集成到修改参数的ResNet18模型中,构建SE-ResNet18模型,其在2.75 MB模型权重大小的基础上平均准确率达到0.9404,表明该算法在德州驴声音分类的准确率和模型轻量化方面具有良好性能。Sound is closely related to healthy animal husbandry,but due to the difficulty of data collection and other reasons,there are still significant challenges in studying animal sound.In this paper,we take the unique Dezhou donkey in China as the research object,and propose a Dezhou donkey sound classification algorithm based on the SE-ResNet18 model that integrates channel attention mechanism.Firstly,spectral subtraction and single parameter dual threshold endpoint detection method are used to preprocess the collected sound signals of donkeys in three different states,and 13 dimensional Mel frequency cepstral coefficients are extracted as features.Secondly,integrate the SE module into the ResNet18 model with modified parameters to construct the SE-ResNet18 model.SE-ResNet18,with an average accuracy of 0.9404 and a model weight size of 2.75 MB,demonstrates good performance in both accuracy and model lightweighting for Dezhou donkey sound classification.

关 键 词：德州驴 声音信号处理 梅尔倒谱系数 ResNet18 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]