Line spectrum extraction of underwater ship engine sound using graph convolutional network

Line spectrum information is crucial for detecting underwater ship engine sounds in passive sonar systems. However, most studies rely on conventional methods or image processing techniques, which are often error-prone. While recent deep-learning approaches show promise, they require precise labels during training, limiting their practicality. To overcome this limitation, we propose a line spectrum extraction method using a graph convolutional network (GCN) that captures correlations between line spectra. The proposed method divides the spectrogram into image patches and represents these patches as a graph, including their connections. Subsequently, a GCN is trained using weak labels that contain significant errors to extract the line spectra. Experiments using publicly available underwater acoustic data demonstrate that the GCN trained with weak labels accurately extracts line spectra while capturing their correlations. Additionally, the proposed method shows superior performance compared to conventional approaches.