Ship-radiated noise separation via auditory scene analysis-guided deep learning

Abstract

Ship-radiated noise is an important acoustic source in passive underwater monitoring, but marine recordings often contain overlapping components from multiple sources, which can hinder target-related analysis. To address this problem, this paper develops a ship-radiated noise separation method that integrates auditory scene analysis with deep learning. Following the computational auditory scene analysis framework, the proposed method consists of auditory segmentation and auditory grouping. In the segmentation stage, a Gammatone filter bank provides time-frequency decomposition, and Dense-UNet extracts deep features and estimates separation masks from the resulting time-frequency representation. In the auditory grouping stage, the separation masks estimated by the network are first mapped from the Gammatone frequency-band domain onto the linear frequency axis. These masks are then applied to the complex spectrogram of the original mixed signal, which helps reduce reconstruction distortion when converting the separated auditory-band representations back to the full-band waveform. The proposed method was evaluated on the ShipsEar and DeepShip datasets. Under the tested conditions, the method consistently achieved competitive performance across multiple objective metrics compared to other commonly used models. Further analysis also revealed that the method exhibits enhanced capability in recovering low-frequency line-spectrum structures, which are crucial for ship-radiated noise characterization.