Page 129 - 应用声学2019年第4期

第 38 卷 第 4 期 Vol. 38, No. 4
2019 年 7 月 Journal of Applied Acoustics July, 2019

⋄ 李启虎院士八十华诞学术论文 ⋄

水下目标多模态深度学习分类识别研究

曾 赛 1,2† 杜选民 1,2

(1 水声对抗技术重点实验室 上海 201108)
(2 上海船舶电子设备研究所 上海 201108)

摘要 水下目标的分类识别对于水声探测具有重要意义。该文提出一种水下目标多模态深度学习分类识别方
法。针对水声信号的一维时域模态和二维频域模态特征建立一种多模态特征融合的深度学习结构，结合长短
时记忆网络和卷积神经网络的优点，对一维时域信号和二维频谱信号分别进行并行处理，对输出进行典型相
关分析，形成特征融合表示，并利用相邻帧的相关性进行参数优化。利用实测水声信号对算法进行了验证。结
果表明：提出的算法对于水下目标识别的精度有显著的提高。
关键词 水下目标识别，长短时记忆网络，卷积神经网络，典型相关分析
中图法分类号: TP391.4 文献标识码: A 文章编号: 1000-310X(2019)04-0589-07
DOI: 10.11684/j.issn.1000-310X.2019.04.016

Multimodal underwater target recognition method based on deep learning

ZENG Sai 1,2 DU Xuanmin 1,2

(1 Science and Technology on Underwater Acoustic Antagonizing Laboratory, Shanghai 201108, China)
(2 Shanghai Marine Electronic Equipment Research Institute, Shanghai 201108, China)

Abstract Underwater target recognition has great significance for underwater acoustic detection. The mul-
timodal underwater target recognition method was proposed based on deep learning. Due to the time domain
features and frequency domain features, a multimodal structure was proposed to incorporate the long short-
term memory neural network and convolution neural network. The time domain modal and frequency domain
modal were processed respectively, the output of those networks was generated as feature fusion by canonical
correlation analysis method. The temporal coherence of adjacent signal frame was utilized to improve the
recognition accuracy. The experiments were implemented based on measured underwater acoustic signal. The
results show that the proposed method improves the accuracy of underwater target recognition significantly.
Key words Underwater target recognition, Long short-term memory neural network, Convolution neural
network, Canonical correlation analysis

2019-02-02 收稿; 2019-05-07 定稿
作者简介: 曾赛 (1989- ), 男, 湖北荆州人, 博士研究生, 研究方向: 水下目标特性、水下目标探测、水声信号与信息处理。
† 通讯作者 E-mail: sharememezeng@126.com