A parallel convolutional neural network-transformer model for underwater target recognition based on multimodal feature learning

Underwater acoustic target recognition is a hot research issue with a wide range of applications. The variable ocean environment and evolving underwater moving target noise reduction techniques greatly complicate the recognition task. Traditional recognition methods are difficult to obtain practical characterization features and robust recognition results due to the singular input features and the limitation of the network backbone. Therefore, We propose a parallel convolutional neural network (CNN)-Transformer model based on multimodal feature learning for underwater target recognition. The CNN module extracts deep features from the Mel-Frequency Cepstral Coefficients (MFCCs). The Transformer captures global information in the original time-domain signal. The two single-modal features are combined by an adaptive feature fusion module to construct joint features for target recognition. The effectiveness of the proposed method was verified in the Ships-Ear dataset, and the average accuracy of classification reached 98.58%. The experimental results show that our model works better than classical methods.