Underwater Acoustic Target Recognition Based on One-Dimensional Convolutional Neural Network with Attention Mechanism

Abstract

To address the issues of complex network parameters and high computational costs in deep learning-based underwater acoustic target recognition models, this study proposed a lightweight one-dimensional convolutional neural network with an attention mechanism for underwater acoustic target recognition. First, during the feature extraction stage, spectral, Mel-spectrogram, chroma, spectral contrast, and tonal features were selected and reconstructed into a fused one-dimensional hybrid feature. Next, the hybrid feature was processed by a multi-scale residual convolution(MRC) module to enhance feature representation across different scales. Simultaneously, a convolutional block attention module(CBAM) was introduced to adaptively adjust feature weights through channel and spatial attention modules, improving the model’s focus on critical regions. Experimental results show that the proposed model achieves an average recognition accuracy of 98.58% on the ShipsEar dataset, demonstrating excellent classification performance and significantly reducing computational complexity.