基于深度学习的低空声目标识别  

Low Altitude Acoustic Target Recognition Based on Deep Learning Technology

作　　者：王显云 王志峰[1] 黄山 WANG Xianyun;WANG Zhifeng;HUANG Shan(NO,3 Research Institute of China Electronics Technology Group Corporation,Beijing 100015,China)

机构地区：[1]中国电子科技集团公司第三研究所,北京100015

出　　处：《电声技术》2022年第3期67-70,74,共5页Audio Engineering

摘　　要：本文提出采用人耳听觉特征和深度神经网络(Deep Neural Network,DNN)相结合的方式对低空飞行目标进行分类。首先,以不同目标的梅尔频率谱(Mel-Frequency Cepstrum Coefficients,MFCC)和伽玛通功率谱(Gammatone Filterbank spectra,GF)为静态特征,并以它们的差分谱作为动态特征;其次,利用谐波处理技术获得具有谐波保护的上述静态特征和动态特征;最后,将上述特征进行组合,作为深度神经网络的输入参数进行网络训练,来进行不同低空声目标的鉴别。试验结果表明,基于深度学习的方法在低空飞行目标识别方面可以取得较好的识别效果。Based on the human auditory features and deep neural network, a passive acoustical recognition technique about low altitude targets is studied in this paper. Firstly, the Mel-Frequency Cepstrum Coefficients(MFCC) and Gammatone Filterbank spectra(GF) of different targets are viewed as static features, and their differential spectra are viewed as dynamic features. Then, the method of harmonic preservation is used to generate the abovementioned acoustic features with harmonic preservation. Finally, the extracted features are viewed as the input parameters of the deep neural network for identifying different targets. The experimental results show that the method based on deep learning can achieve good classification in low altitude acoustic target recognition.

关 键 词：低空声目标识别 深度神经网络(DNN) 梅尔频率倒谱系数(MFCC) 伽玛通功率谱(GF) 

分 类 号：TN912.1[电子电信—通信与信息系统]