摘要
基于深度学习的鸟鸣声识别是当前研究的热点,现有基于语谱图的识别方法无法提取帧间的时序信息,文章提出了一种基于C-LSTM(CNN-Long Short-Term Memory)的鸟鸣声识别方法,该方法以梅尔语谱图为输入,通过CNN提取谱图特征后,输入到LSTM模型中,进一步提取不同帧之间的时序特征,基于该特征实现鸟鸣声的分类。选择Xeno-Canto中的5种鸟类作为研究对象,对比了VGG16模型和C-LSTM模型的平均识别准确率(Mean Average Precision,MAP)值。结果表明,以VGG16和C-LSTM作为识别模型时,测试集的MAP值分别为0.8628和0.9147,文章提出模型的MAP提升5.19%。说明文章提出的C-LSTM更适合于鸟类物种识别,具有更高的识别性能。
Birdsong recognition based on deep learning is the focus of research at present.The existing recognition methods based on spectrogram cannot extract the timing feature between signal frames.This paper presents a method of bird song recognition based on C-LSTM(CNN-Long Short-Term Memory)model.In this method,Mel Cepstrum is used as the input.After the spectral features are extracted by CNN,they are fed into the LSTM model to extract the sequential characteristics between different frames.Based on this feature,we realized the classification of birdsong.We selected five bird species in Xeno-Canto as the research objects,and compared the Mean Average Precision(MAP)of VGG16 model and C-LSTM model.The results show that the MAP values of VGG16 and C-LSTM are 0.8628 and 0.9147,respectively.The MAP of C-LSTM model is improved by 5.19%.This shows that the C-LSTM proposed in this paper is more suitable for bird species identification and has higher recognition performance.
作者
邢照亮
吴伟银
张正晓
陈麒麟
倪东明
XING Zhaoliang;WU Weiyin;ZHANG Zhengxiao;CHEN Qilin;NI Dongming
出处
《科技创新与应用》
2021年第15期15-18,共4页
Technology Innovation and Application
基金
国家电网公司科技项目(编号:SGGR0000WLJS1801082)
北京林业大学教育教学研究项目(编号:BJFU2019JY051)。
关键词
鸟鸣声识别
时序特征
长短时记忆网络
birdsong recognition
timing features
LSTM(Long Short-Term Memory)
