DSCANet: Underwater Acoustic Target Recognition Using the Depthwise Separable Convolutional Attention Module

Abstract The classification and recognition of underwater acoustic targets is a critical support technology for underwater acoustic information countermeasures. The extraction and classification of features of the underwater target is its research focus. Researchers have done an in-depth study from multiple perspectives. However, because the underwater acoustic signal is affected by ambient noise and operating conditions of different targets, the signal-to-noise ratio of the underwater acoustic signal is very low, the components of underwater acoustic signal are complicated, and specific signal pre-processing is usually necessary, such as signal enhancement and signal decomposition. In existing methods, the main focus is on extracting and classifying underwater acoustic signal features after multi-step preprocessing. However, these methods overlook the integration of feature extraction and classification, leading to certain drawbacks. To overcome these limitations, we propose a new model called Depthwise Separable Convolutional Attention (DSCA). This model can better recognize and classify underwater acoustic targets. We also use multiple DSCAs to construct a neural network, which we called DSCANet. The DSCANet can be used for classification and recognition of underwater acoustic targets. The structure of the DSCANet is unified and simple, and no specific pre-processing of the underwater acoustic signal is necessary. Spatial structure features are extracted from bottom layers of the DSCANet, and semantic features are extracted from top layers of the DSCANet, and then classifying them. The DSCANet integrates feature extraction and target classification for underwater acoustic targets. The DSCANet is trained and validated on ShipsEar, an open dataset, the classification accuracy reaches 93%, which is the highest in contrast experiment.