Multi-Features Fusion for Underwater Acoustic Target Recognition based on Convolution Recurrent Neural Networks

Abstract

The underwater acoustic target recognition task has raised great concern in recent years, while it remains a huge challenge, because of the difficulties of accessing clean and labeled underwater acoustic data, the complicacy of sound wave transmission in the epicontinental sea, and the interference from surrounding sound sources. In this paper, we proposed a deep learning-based underwater acoustic recognition method, called FCRN, in which a convolution recurrent neural network (CRN) is utilized to process the multi-dimensional time-frequency domain features after fusion. Our proposed method is evaluated and discussed on the measured underwater acoustic dataset ShipsEar. The experimental results have indicated that the proposed FCRN model is effective in underwater acoustic target recognition task, achieving an accuracy ratio is 96.67%.