嵌入注意力机制的卷积神经网络水声目标识别

针对低信噪比水声目标单一特征识别率低，稳健性差的问题，提出一种基于注意力机制和多尺度残差卷积神经网络（Multi-scale Residual CNN with Attention，MR-CNN-A）进行特征融合的识别方法。该方法根据多尺度卷积核与特征图形成多分辨率分析关系，并以此通过注意力机制实现优势特征权值提取与融合，从而提高模型在文中水声数据集上提取目标噪声特征和分类识别的稳健性与抗噪能力。开展了4类舰船噪声和海洋环境噪声的识别试验、水下和水面自主式水下航行器的识别试验，以及不同信噪比条件下目标噪声的识别试验。结果表明：对于文中所涉及的水声目标噪声和人工高斯白噪声干扰，该网络模型识别正确率明显高于支持矢量机与简单卷积神经网络，且对高斯白噪声的抑制能力远强于支持矢量机与简单卷积神经网络，稳健性好，模型复杂度小。

Underwater target recognition by CNN with attention mechanism

Because of low recognition rate and poor robustness to a single feature of underwater acoustic target with low signal to noise ratio (SNR), a feature fusion recognition method based on attention mechanism and multi-scale residual convolution neural network (named MR-CNN-A network) is proposed. According to the multi-resolution analysis relationship formed by multi-scale convolution kernel and feature map, and by using the attention mechanism to extract and fuse the dominant feature weights, this method can enhance the robustness and noise immunity of the model on the underwater acoustic data set. Four types of ship noises and the marine ambient noise recognition tests, underwater and surface autonomous underwater vehicle (AUV) recognition tests, and the recognition tests of target noises under different signal to noise ratios are carried out. The results show that: for the interferences of underwater acoustic target noise and artificial Gaussian white noise involved in this paper, the recognition accuracy of the network model is obviously higher than that of support vector machine and simple convolutional neural network. The model is more powerful in suppressing Gaussian white noise than support vector machine and simple convolutional neural network, and has good robustness and low complexity.