水下目标多模态深度学习分类识别研究  

Multimodal underwater target recognition method based on deep learning

作　　者：曾赛 杜选民[1,2] ZENG Sai;DU Xuanmin(Science and Technology on Underwater Acoustic Antagonizing Laboratory,Shanghai 201108,China;Shanghai Marine Electronic Equipment Research Institute,Shanghai 201108,China)

机构地区：[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

出　　处：《应用声学》2019年第4期589-595,共7页Journal of Applied Acoustics

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。Underwater target recognition has great significance for underwater acoustic detection. The multimodal 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 shortterm 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.

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 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]