改进的卷积神经网络实现端到端的水下目标自动识别

由于水声信号的高度复杂性，基于特征工程的传统水下目标识别方法表现欠佳。基于深度学习模型的水下目标识别方法可有效减少由于特征提取过程带来的水声信号信息损失，进而提高水下目标识别效果。本文提出一种适用于水下目标识别场景的卷积神经网络结构，即在卷积模块化设计中引入卷积核为1的卷积层，更大程度地保留水声信号局部特征，且降低模型的复杂程度；同时，以全局平均池化层替代全连接层的方式构造基于特征图对应的特征向量主导分类结果的网络结构，使结果更具可解释性，且减少训练参数降低过拟合风险。实验结果表明该方法得到的水下目标识别准确率（91.7%）要优于基于传统卷积神经网络（69.8%）和基于高阶统计量特征的传统方法识别表现（85%）。这说明本文提出的模型能更好保留水声信号的时域结构，进而提高分类识别效果。 

End to End Underwater Targets Recognition Using the Modified Convolutional Neural Network

Traditional feature-based underwater target recognition methods perform poorly due to the high complexity of underwater acoustic signals. Advanced recognition methods based on the deep learning model can effectively reduce the information loss caused by the feature extraction, thereby improving the classification performance. In this paper, we proposed a convolutional neural network (CNN) model suitable for the underwater targets recognition scenario, which introduced a one-dimension Convolution layer with the kernel of 1 in the convolution module to preserve the local characteristics of underwater acoustic signals and reduce the complexity of the model; meanwhile, replaced the fully connected layer with a global average pooling (GAP) layer which outputted the interpretable results based on the feature vector corresponding to feature map and reduced the training parameters to prevent overfitting. The results showed that the modified CNN model achieved a classification accuracy of 91.7%, compared with the classification method based on conventional CNN which obtained 69.8% and features of higher-order statistics (HOS) which obtained 85%. It is concluded that the proposed method can better preserve the time-domain structure of underwater acoustic signals, furthermore improving the classification performance.