联合线谱增强与深度神经网络的水声目标识别

Underwater acoustic target recognition using line enhancement and deep neural network

为强化水声目标特征,提高使用深度神经网络识别水声目标的准确率,提出了一种联合线谱增强与深度神经网络的水声目标识别方法。该方法采用窄带信息增强,将自适应线谱增强滤波器与VGGish神经网络级联,水声信号经过线谱增强后输入网络提取深度特征,之后使用分类器分类。使用实测水声数据集进行测试,对网络提取的水声数据的深度特征集进行主成分分析并降维,使高维深度特征可视化,结果表明线谱增强后得到的深度特征集的紧致性明显提高。该方法在测试数据集上能够实现94.83%的识别准确率,与未进行线谱增强的情况相比提升了5.48%,同时在低信噪比情况下稳定性更好。

To enhance the features of underwater acoustic target signals,and improve the performance of underwater acoustic target recognition based on deep neural network,a target recognition method using line enhancement and deep neural network is proposed.This method focuses on the narrowband information enhancement and sets an adaptive line enhancement filter at the front end of the VGGish network.The signals are processed by the line enhancement filter and input into the network to extract deep features,and then these features are classified by a classifier.The effectiveness of the method is verified by the actual underwater acoustic dataset.Principal component analysis is performed on the deep feature set of the underwater acoustic signals,and the results show that the compactness of the deep feature set obtained after line enhancement is significantly improved.The proposed method can obtain a recognition accuracy of 94.83%on the test dataset,which is improved by 5.48%compared to the case without line enhancement,and it is also more robust under the condition of low signal-to-noise ratio.