Conditional GAN-Based Whistle Generation for Underwater Biomimetic Acoustic Communication

초록

This paper proposes a conditional whistle generation technique that can selectively generate whistles of desired shapes for underwater biomimetic acoustic communication. The proposed technique performs clustering based on whistle shapes and designs a conditional generative adversarial network (cGAN) structure that can selectively generate whistles of specific shapes by utilizing the labels of each cluster as condition information. In addition, to enable highresolution whistle generation, a resolution progressive growing method was employed, wherein only three resolution stages32 × 32,128 × 128, and 512 × 512-were selectively applied to reduce computational complexity while maintaining highresolution generation capability. Performance evaluation results showed that the proposed model recorded an average Fréchet inception distance(FID) score approximately 330 points lower than conditional deep convolutional generative adversarial network(cDCGAN) in all clusters. Visual analysis also confirmed that the proposed model generated clear whistle images reflecting the shape of each cluster. © 2025 IEEE.

키워드