Sonar Waveform Synthesis Based on Mel Generative Adversarial Networks with Self-Attention Mechanisms

Le résumé fourni par la source

Considering the complicated process of traditional sonar waveform synthesis method, in this study, we apply models from the speech domain to the underwater acoustic domain based on deep learning algorithms using a transfer learning approach. The core of this study lies in the development of a transfer learning framework based on the generative adversarial network (GAN) model, which effectively adapts to the special characteristics of sonar waveforms by fine-tuning the pretrained MelGAN model and the self-attention mechanism. Because sonar waveform cannot be accurately assessed by auditory assessments, we adopted Fréchet Audio Distance (FAD) as an objective audio assessment metric. Experiments based on the DeepShip dataset show that the proposed model effectively synthesizes a sonar waveform.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.