Abstract:
Automatic classification of marine vessels has gained significant interest due to its growing non-military uses such as management of maritime traffic, controlled fishing, and marine environment protection. An AI-powered, reliable maritime traffic monitoring system that can identify the source of noise in underwater environments is the need of the hour. The work at hand proposes an extended SRN-AST (Ship-Related-Noise based Audio Spectrogram Transformer) and SRR-SSAST (Ship-Related-Noise based Self-Supervised Audio Spectrogram Transformer) model, which leverages attention mechanisms to capture both local and global features from spectrograms. The extended model uses fewer parameters, supports parallel processing, and outperforms traditional deep models, demonstrating the effectiveness of transformers in underwater vessel detection, recognition, and identification. The model has been tested on DeepShip, a publicly available underwater acoustic benchmark dataset. We achieve an accuracy of 97% with the enhanced SRN-AST algorithm and an accuracy of 99.5% with the enhanced SRN-SSAST algorithm. This approach eliminates the need for labeled pre-training and aims to excel in real-world underwater signal classification tasks.
