基于多特征融合的蛋鸡发声识别方法研究  

Recognition Method of Laying Hens’Vocalizations Based on Multi-feature Fusion

作　　者：余礼根[1,2] 杜天天 于沁杨 刘同海 孟蕊 李奇峰 YU Ligen;DU Tiantian;YU Qinyang;LIU Tonghai;MENG Rui;LI Qifeng(Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;National Engineering Research Center for Information Technology in Agriculture,Beijing 100097,China;College of Computer and Information Engineering,Tianjin Agricultural University,Tianjin 300384,China)

机构地区：[1]北京市农林科学院信息技术研究中心,北京100097 [2]国家农业信息化工程技术研究中心,北京100097 [3]天津农学院计算机与信息工程学院,天津300384

出　　处：《农业机械学报》2022年第3期259-265,共7页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(31402113);北京市农林科学院青年基金项目(QNJJ201913);北京市农林科学院创新能力建设专项(KJCX20211007);广东省重点领域研发计划项目(2019B020217002)。

摘　　要：为更好地利用音频进行畜禽发声分类,进一步提高识别准确率,提出了一种基于多特征融合的蛋鸡发声识别方法。以栖架式养殖模式下蛋鸡的产蛋声、鸣唱声、饲喂声、尖叫声典型音频为研究对象,提取梅尔频谱系数、短时过零率、共振峰及其一阶差分作为融合特征参量,构建基于遗传算法优化BP神经网络的蛋鸡发声分类识别模型。结果表明,本文方法对蛋鸡产蛋声、鸣唱声、饲喂声和尖叫声的平均识别准确率为91.9%,识别的精确度分别为90.2%、93.0%、93.3%、92.2%,平均精确度达到92.2%;识别的灵敏度为94.9%、90.0%、89.4%、91.8%,平均灵敏度达到91.5%。研究表明,基于多特征融合的蛋鸡发声识别方法具有较好的识别灵敏度和精确度,可为蛋鸡发声语义解析与自动判别提供参考。Vocalization is a direct expression of poultry's rich body information,physiological characteristics,stress response and health status,which can be used to characterize emotional health changes,physiological growth feedback,and feeding regulation with the advantages of non-invasive,non-stress and continuous monitoring.In order to make better use of audio multi-dimensional features to classify poultry vocalizations,a recognition method for laying hens'vocalizations based on multi-feature fusion was proposed.Typical calls of laying hens such as egg laying,singing,feeding and screeching in perching system were collected and analyzed,the Mel frequency cestrum coefficient,short-time zero-crossing rate,formants and first-order difference were computed by Matlab software.The classification and recognition models of laying hens'vocalizations were established based on genetic algorithm optimized BP neural network according to the multi-feature fusion.The results showed that the average recognition rate by this method for laying hens'sounds of egg laying,singing,feeding and screeching was 91.9%,and the accuracies were 90.2%,93.0%,93.3%and 92.2%,respectively;and their sensitivities were 94.9%,90.0%,89.4%and 91.8%,respectively.The average accuracy and sensitivity were 92.2%and 91.5%,respectively.It was found that this recognition method of laying hens'vocalizations based on multi-feature fusion had a higher classification accuracy and sensitivity,which could be used for automatic discrimination and classification for different livestock and poultry sounds.

关 键 词：蛋鸡发声 栖架养殖 多特征融合 分类识别 梅尔频谱系数 短时过零率 

分 类 号：S831.4[农业科学—畜牧学]