CNN-Transformer Feature Aggregation Networks for Underwater Acoustic Target Recognition

This paper introduces the CAT model, achieving 98.9% accuracy on the ShipsEar dataset by combining CNN and ViT with a fusion and separation mechanism for underwater acoustic signal processing. The model captures both short-term and long-term dependencies through integrated feature extraction and utilizes multi-spectral preprocessing to enhance performance. CAT shows potential for applications in speech recognition and environmental sound classification, with future work focusing on real-time monitoring and additional input features.