Underwater target recognition using a lightweight asymmetric convolutional neural network

Underwater acoustic target recognition(UATR) is a critical research issue in marine acoustics. Nonetheless, due to the interference from irregular noise and variable channel transmission environment, traditional recognition methods for underwater targets have difficulty adapting to complex and changeable ocean environments. The feature extraction method combined time-frequency spectrograms with Convolutional Neural Networks(CNN) can effectively describe the differences between various targets. However, many existing CNNs are not suitable for applying to embedded devices because of their high computational costs. To this end, we propose a lightweight network based on an asymmetric convolutional neural network (LW-A-CNN) for UATR. LW-A-CNN can capture more stable low-frequency line spectrum features and maintain its lightweight by employing asymmetric convolutions to balance accuracy and efficiency. Experiments on the shipsear dataset show that LW-A-CNN achieves the highest recognition accuracy of 98.9% compared to other state-of-the-art deep learning methods and significantly decreases model parameter size. Additionally, LW-A-CNN demonstrates robust performance against interference.