A Novel Underwater Acoustic Signal Denoising Model Based on Complex Convolution Dual-branch Multi-scale Attention Network

Abstract

With the rapid advancement of underwater target stealth technology, the development of efficient denoising and signal restoration techniques for ultra-low signal-to-noise ratio (SNR) underwater acoustic target signals has become an urgent research priority. To address this challenge, this paper proposes an Underwater Acoustic Signal Denoising Model Based on Complex Convolution Dual-branch Multi-scale Attention Network (CCDB-MSANet). Specifically, CCDB-MSANet introduces a complex spatial coordinate convolution method, enhancing the network's ability to perceive spatial positions and understand the spatial features of time-frequency spectrograms. Next, a dual-branch multi-scale feature enhancement network is designed to comprehensively amplify the amplitude, phase, and energy-related features of the complex time-frequency spectrum of the target signal. Additionally, a complex spatial sub-pixel block is developed to accurately map the complex spectral features from low-dimensional to high-dimensional spaces, minimizing target information loss and reducing artifacts. Finally, a non-linear mask block is designed to compensate for the model's non-linear learning capabilities. Experimental results demonstrate that the proposed CCDB-MSANet significantly outperforms traditional signal analysis and deep learning-based methods in denoising ultra-low SNR underwater acoustic signals. More importantly, the model achieves equally outstanding performance on unknown underwater target datasets across multiple sea environments, showcasing its potential as a robust system for real-world ultra-low SNR underwater acoustic signal denoising applications.