InfoGAN-Enhanced Underwater Acoustic Target Recognition Method Based on Deep Learning

Abstract

Aiming at the small sample problem of underwater acoustic target recognition (UATR) task, this paper proposes a method which uses InfoGAN to generate underwater acoustic target data for UATR. In this method, the underwater acoustic target audio data is transformed into time-frequency image data by short-time Fourier transform, and then input the time-frequency image data into InfoGAN to generate underwater acoustic target data. Finally, the method establishes the IG_DCNN recognition model and the real data samples are mixed with input training. To test and verify the feasibility and effectiveness of the InfoGAN-enhanced UATR method, the original dataset was input into CNN to establish a comparative model. The experimental reveal that InfoGAN can generate diverse underwater acoustic target data samples, and the IG_DCNN model constructed by the InfoGAN-enhanced UATR method proposed in this paper has better network performance and recognition accuracy.