Underwater bubble plume image generative model based on noise prior and multi conditional labels

Abstract

The characteristics of underwater bubble plume dispersion behavior are important to assess and manage potential risks of underwater gas pipelines. In order to generate sufficient image samples for underwater bubble plume research based on artificial intelligence, we try to design an optimized generative adversarial network. In this paper, a novel underwater bubble plume image generative model that combines noise prior and multi conditional labels is proposed and evaluated using different testing dataset. In the proposed framework, the noise prior which contains the image attributes and the category features are trained with the help of VAEs, prior noise and multi conditional labels are utilized to construct the generative and discriminative models. Finally, 3 types of the underwater bubble plume images generated from the experiment is contrasted and evaluated. The experimental results show that，compared with the existing methods text2img and CGANs，the FID values of the generated images are reduced by 97.4% and 22.8% on SUIM dataset, and the FID values are reduced by 97.4% and 22.8% on BUBBLE dataset, our proposed framework achieves satisfactory and promising performance.