CA_MobileNetV2 for Underwater Acoustic Target Recognition.

In response to the hardware limitation and realtime requirement in the underwater acoustic target recognition, this paper proposes a coordinate attention-based end-to-end MobileNet (CA_MobileNetV2) with fewer parameters and high recognition accuracy. The acoustic signal is mapped into LOFAR images and input into the backbone MobileNetV2. We utilize depth-wise separable convolution to reduce model complexity. However, insufficient feature extraction of lightweight models may lead to accuracy loss. CA_MobileNetV2 embeds Bottleneck's acoustical time-frequency characterized Coordinate Attention (CA) mechanism. This mechanism captures long-term temporal dependencies and preserves the precise positional relationships in the frequency domain to represent the target signal. Additionally, the cosine annealing algorithm is used to enhance the convergence speed of the model during training, preventing the model from getting stuck in local optima and reducing the model's accuracy. Experimental results demonstrate that the CA_MobileNetV2 reduces the parameter size by 80.05% compared to ResNet18 and achieves an accuracy of 98.16% with a more pronounced inter-class gap in underwater acoustic target recognition.