用于水声目标识别的注意力机制特征增强  

A feature attention enhancement network for underwater acoustic classification

作　　者：任佳威 谢源 徐及[1] 颜永红[1] REN Jiawei;XIE Yuan;XU Ji;YAN Yonghong(Key Laboratory of Speech Acoustics and Content Understanding,Institute of Acoustics,Chinese Academy of Sciences,Beijing,100190;University of Chinese Academy of Sciences,Beijing,100049)

机构地区：[1]中国科学院声学研究所语言声学与内容理解重点实验室,北京100190 [2]中国科学院大学,北京100049

出　　处：《声学学报》2024年第4期664-670,共7页Acta Acustica

基　　金：中国科学院声学研究所自导项目(mbdx202107)资助。

摘　　要：提出了一种基于注意力机制的特征增强网络(FEAN)用于水声目标的识别,该方法以水声目标辐射噪声信号的线谱信息和调制信息为基础,采用可学习的注意力模块,一方面依照后端的分类任务,对特征进行线性的自适应滤波操作,保留对分类任务有效的信息,并且滤除干扰;另一方面利用线谱信息在时域上的相干性,调制信息在频域上的相关性来优化时频谱图特征。在DeepShip和Shipsear公开数据集上的实验结果表明,相比于以传统的短时傅里叶变换(STFT)和Fbank特征作为输入的识别模型,FEAN在两个数据集上分别获得了4.5%和1.2%的识别准确率提升。A feature attention enhancement network(FEAN)is proposed for underwater acoustic classification.Based on the line spectrum information and modulation information of radiated noise signals from underwater acoustic targets,the FEAN employs a learnable attention module that performs linear adaptive filtering on features according to the classification task,retaining information relevant to the classification task while filtering out interference.Additionally,it optimizes time-frequency spectrogram features by utilizing the coherence of line spectrum information in the time domain and the correlation of modulation information in the frequency domain.The experimental results on the DeepShip and Shipsear public datasets show that compared with the recognition models with the traditional short-time Fourier transform(STFT)and Fbank features as input,the FEAN has achieved a 4.5%and 1.2%improvement in recognition accuracy on the two datasets,respectively.

关 键 词：水声目标识别 深度学习 注意力机制 特征增强 

分 类 号：TB566[交通运输工程—水声工程]