A Two-Stream Network for Underwater Acoustic Target Classification

Abstract

Due to the complex marine environment, underwater acoustic signals may lose some feature information in the process of transmission, resulting in limited classification accuracy. In order to achieve higher classification accuracy, we propose a novel network structure to deal with different features of underwater acoustic signals. In the processing of underwater acoustic signal, the accuracy of Visual Geometry Group (VGG) is 2.3% higher than ResNet50, and the accuracy of Gate Recurrent Unit (GRU) is 1.1% higher than Long Short-Term Memory(LSTM) and 4.2% higher than Recurrent neural network (RNN). The proposed method consists of two parts: (1) MFCCNet: A GRU based network for training features from MFCC. (2) SpecNet: Using VGG to process features from spectrogram. The two parts are connected by a fully connected layer for the final output. The integration of SpecNet and MFCCNet promotes the whole network to learn deeper features. Experiments show that our method achieves 98.8% accuracy in the actual data set of civil ships.