摘要
在水声目标识别领域,现有识别方法主要采用基于时域或时频域技术,对环境噪声和干扰非常敏感,尤其在复杂多路径水声环境中,信号容易受到干扰,且仅依赖时域和时频域的特征难以准确描述复杂或相似目标的关键属性。对此,为进一步挖掘水声数据信息,并充分利用时域与时频域特征,提出了基于多模态特征交叉融合的水声目标识别网络。首先,利用一维序列特征提取模块和以ResNet为主干网络的二维图像特征提取模块对水声数据进行粗提取。其次,为准确捕捉时域中时序信息,设计了一种时序注意力模块,实现对不同时间步的特征提取。接着,通过提出的特征融合模块实现多模态特征的融合,提升网络对多模态特征的提取能力,增强网络的特征表达能力。最后,在DeepShip数据集上进行验证,提出的水声目标识别网络取得98.80%的准确率,优于其他方法,证实了所提方法的有效性。
In the field of underwater acoustic target recognition,existing methods primarily utilize time-domain or time-frequency domain techniques,which are highly sensitive to environmental noise and interference.This sensitivity is particularly problematic in complex multipath underwater environments,where signals are prone to disturbances,and reliance solely on time-domain and time-frequency domain features often fails to accurately describe the critical attributes of complex or similar targets.To address this,we propose a Multimodal Cross-Feature Fusion Network(MCFNet)for underwater acoustic target recognition.Firstly,coarse extraction of underwater acoustic data is performed using a one-dimensional sequence feature extraction module and a two-dimensional image feature extraction module based on ResNet.Secondly,to accurately capture temporal information in the time domain,a Temporal Attention Module(TAM)is designed to extract features from different time steps.Subsequently,a proposed Cross-Attention Feature Fusion Module(CAFM)integrates multimodal features,enhancing the network's ability to extract and express features.Finally,MCFNet is validated on the DeepShip dataset,achieving an accuracy of 98.80%,which surpasses other methods.The results confirmed the effectiveness of the proposed method.
作者
包宇航
郑成伟
魏长赟
BAO Yuhang;ZHENG Chengwei;WEI Changyun(College of Mechanical and Electrical Engineering,Hohai University,Changzhou 213200,China)
出处
《无人系统技术》
2024年第4期55-65,共11页
Unmanned Systems Technology
基金
国家自然科学基金(52371275)。
关键词
水声目标识别
特征提取
多模态特征融合
信息交互
深度学习
目标分类
Underwater Acoustic Target Recognition
Feature Extraction
Multimodal Feature Fusion
Information Interaction
Deep Learning
Target Classification
